I. The music industry is trying to work out new business models in AI amid lawsuits

The music industry is still battling with AI companies over copyright cases with regard to using data in model training. But it is also thinking about how to make the most of the situation and earn money. Last week, Suno released a new model, which is trained from scratch with licensed data from Warner Music Group, BMG, and Believe.

On the other hand, ElevenLabs signed a partnership with Universal Music Group, which is in an active legal battle with Suno. The voice AI company also released a new v2.5 model and enabled lossless downloads. We will see more “industry-approved” models in the future, and their sound signature will depend on the catalog of data from labels and other sources. Some models might be great at background scoring, and some might be great for ads.

As Musixmatch co-president Rio Caraeff told Week in Voice AI in an interview, we will see more collaborations as it’s monetarily beneficial for platforms, rights holders, and model makers to work together.

II. India’s Global Fintech Fest becomes a voice AI fest

The finance industry is one of the best targets for voice AI companies. This tweet possibly sums it all up.

There were new partnerships and products announced, including Sarvam’s partnership with Mahindra Finance, Cashfree’s voice-powered payment recovery stack, and Arrowhead’s own pipeline for BFSI clients. In the coming months, we might get to know about more deals signed between finance institutions and voice AI companies. Right now, the mood in India around voice AI is buoyant, and startups are taking advantage of that by signing new deals.

In Focus

Apple tries to weave privacy into its new Audio Intelligence features

Apple released two new features with its new Apple Watch in an attempt to indicate that we don’t yet need separate AI gadgets for productivity and accessibility. First is a Live Rewind feature, which transcribes the last 15 seconds of audio, in case you missed anything a person said. The second Siri Recap, a Granola-like meeting notetaker summary without any transcript, along with action items and key points of a conversation.

Both those features sound like always-on features on the face of it. Apple did specify that you can turn them off at any point, and users have full control. But it’s also interesting to understand the processing pipeline of how the company handles privacy for these features when they are on.

Apple Watch 12 and Apple Watch Ultra 4 have a secure enclave within the S11 chip for initial secure processing of audio. For iPhone 16 and above, there is a Secure Enclave in the chip for raw audio processing.

Apple uses a new audio-verified pairing besides Bluetooth to ensure that audio data is sent only between two secure enclaves while keeping the communication encrypted. Backed-up Live Rewind or Siri Recap is encrypted end-to-end.

Live Rewind

Live Rewind works when you double-press the crown on your Apple Watch and shows the last 15 seconds of transcript. Apple Watch keeps a rolling buffer within its chip, so there is no continuous conversation transcript. There is only a backup if you ask Siri a follow-up question or ask it to save it in the Siri app.

Apple said it is also alert people around you with a sound and an on-screen message that you are using the Live Rewind feature:

When activated, an audible chime plays from the speaker on your Apple Watch, even if your Apple Watch is on silent or you have headphones connected. A full display animation and microphone indicator appear on the Apple Watch display, and a distinct double-press gesture is required to activate Live Rewind. This is to alert those nearby that the feature has been activated.

Siri Recap

There have been calls for Apple to buy Granola, but I think for Apple, storing transcripts and all kinds of recording-related data could be privacy trouble in waiting. The company’s new feature is the closest iteration of a meeting notetaker or a conversation capture.

With this feature, Apple doesn’t keep any transcript and also removes financial data, government-assigned identifiers, authentication data, and certain personal identifiers. Personally, I think meeting notetakers should have this toggle for their apps too. Users will get a summary, action items, and highlights from a conversation.

What’s more, users can set a schedule to have this feature on “At work” or turn it off at night.

When the feature is on, Apple runs a conversation detection model on the Apple Watch. Once it detects the conversation, it encrypts the audio and starts transferring it to the iPhone. A model on the Secure Enclave starts transcribing it, and the final product is roughly half of the original transcript, removing tone markers, non-essential language, filler words, and redundancies while preserving the topics.

Then, to further summarize this block of text, it is sent to Apple’s Private Cloud Compute (PCC) for further processing. Apple collates data such as the title of a meeting, if any, in the calendar, rough location like a school, cafe, or a park, and uses it to assign a title to the final summary. Apple uses its own foundation model to generate a summary and key points, and discards the audio and the transcript.

The company said that it doesn’t notify anyone if the Siri Recap feature is on because the summary is “comparable to notes a person might write after a conversation” and there is no saved transcript. Though its support page reads:

While audio is not recorded and stored, and no speakers are attributed, consider those around you where conversations may be private or sensitive.

What does it mean for meeting notetakers?

Like I wrote last week, I don’t think meeting notetakers will be impacted by this. This might work in their favor if people like getting meeting notes and would want advanced features.

A lot of users value transcripts and recordings, which meeting notetakers offer. They also use these transcripts with other AI tools to create custom workflows.

However, meeting notetakers do need to provide some of the privacy protections offered by Apple as optional features, such as private data filtering, secure processing, or discarding transcripts.

How will this change user behavior?

This is a tricky question given people can record conversations all the time, and despite Apple’s filters, sensitive information can still leak through.

There are also hanging legal questions in two-party or all-party consent regions as to how lawmakers will interpret this feature if there is a lawsuit.

I have heard people say “I am not running Granola” when they are at a dinner meeting because they assume the other person is always recording.

Yes, apps like Granola and Circleback are present on Apple Watch, but their scale is a lot less than the number of Apple Watches sold every year. Months from now, we are likely to see stories about people sneakily recording conversations and some sensitive information leaking into summary notes.

Share

Numbers Game

• A survey of 550 decision-makers found that AI usage dropped from 95% to 64% year-over-year.

• India’s Sarvam said that it has handled 325 million voice AI minutes in the preceding 12 months.

• Another Indian startup Navana.ai, said it has handled 100 million voice AI minutes to date.

  ---

Quick Bytes

• China looks into identity rights for citizens to protect them from unsolicited voice cloning

Countries worldwide are looking at preventing deepfakes of people to prevent scams. China is also looking into these issues for its population of over a billion people. The country’s top court is issuing guidelines around how to handle cases related to impersonation or improper usage of voices and identity.

• Yemeni forces were reportedly misled by voice clones

Reports mentioned that Yemeni forces received an audio resembling Lieutenant General Tareq Saleh’s voice giving an order to retreat. The forces accused Iran-backed Houthi fighters of using AI. We will see more usage of AI in warfare and politics to deflect and deceive.

• GPT-Live-1 voice model is now available via API

OpenAI has made its first full duplex model available through API. The company also demoed use cases with its early partners, including Yelp for reservations, Hatch service bookings, Cognition for Devin AI agent, Speak for its AI-powered tutor, HeyGen for its live avatars, and PicsArt to access creative tools.

• LG denies its TVs are listening to you for ad targeting

Last week, GamersNexus reported that hackers can take advantage of LG’s ever-listening capabilities for the wake word. After that, LG defended its stance, saying that its mics are only listening for the assistant hotword and the mics are not meant to record conversations. But as ambient devices around us grow, we will see more concerns around privacy.

Deals Corner

Navana.ai ($4.2 million): Indian startup that builds calling agents, especially for the BFSI industry

• Investors: Ronnie Screwvala (lead), Antler India, Sharad Sanghi, Sandeep Singhal, and Paula Mariwala.

Tellia ($5 million): A voice-first interface for agricultural field teams

• Investors: Revent (lead), Grey Silo Ventures, Jeriko, and Fund F.

Thank you for tuning in. Keep listening.

This newsletter is by Ivan Mehta, a freelance reporter at TechCrunch. It covers AI and technology in voice, audio, and music. Email: voiceaiweek@gmail.com or im@ivanmehta.com

Yesterday, Apple launched a similar feature, but in its own style.

The new iPhone Duo was the star of the show at Apple’s fall event, but there was a section in the Apple Watch presentation that caught my eye the most. In that segment, Apple described its new Audio Intelligence feature, which acts like an always-on meeting notetaker. In my mind, I yelled, “Granola!” Things are a bit more nuanced than that.

Ron Huang, the company’s VP of sensing and connectivity, said that the Apple Watch is a “perfect device to help you stay aware of the sounds around you and keep track of your conversations.”

We will get to the summarizing feature in a second, but Apple also debuted a feature called Live Rewind, which generates a transcription of the past 15 seconds of audio. The demo showed a waitress telling customers about new specials at the restaurant at a fast pace, and one of the customers hits the Digital Crown twice to get a transcript of that. But, like reporter Joanna Stern said, why not just ask her again? There is possibly an accessibility angle, but it would be strange to pull up the last 15 seconds of transcript while missing the current conversation. It is likely that the company keeps 15 seconds of audio/transcript in buffer on rolling basis.

”This is great for when you're trying to recall a book recommendation, a colleague's brilliant idea, or something you didn't hear the first time,” Huang said. If this happens in person, you can ask them again. One potential use case is you listening to a live lecture and missing something, but then again, you might be using something to record it already.

Apple said that this feature doesn’t continue to transcribe in the background, but I assume it has to keep track of the last 15 seconds all the time.

Then there is the Siri Recap feature, which is like any other meeting notetaker. A key difference is that you can’t get a transcript and only get summaries, action items, and takeaways.

There is an argument that this will kill meeting notetakers. I don’t think that is the case in the slightest. First, a lot of people, including journalists like myself, use meeting recorders to obtain the recording or transcript. For AI-powered meeting notetakers, the transcript is a crucial part, as it acts as fodder for prompts and action items that could result in further workflows. Apple doesn’t want to get into any of that. Apple’s new feature could actually push people towards more capable notetakers if needed.

Subscribe now

There are some privacy-focused steps that could separate this feature from meeting notetakers. First, there is no speaker recognition. So users won’t be able to access who said what. Second, Apple said it will remove sensitive information like financial details or government-assigned identifiers. On the face of it, this sounds good. But it is hard to define what might be sensitive to someone, and Apple will need to iterate on its software to make this definition more inclusive.

Third, to produce the summary, Apple processes the audio on the device, creates a transcript, sends a condensed version to the private cloud, and produces the final summary.

While all these steps are good for user protection, sensitive information can still leak in. And Apple is aware of that.

“While audio is not recorded and stored, and no speakers are attributed, consider those around you where conversations might be private or sensitive,” the company writes on a support page.

Apple is setting this up nicely for ever-recording AI wearables that might have fewer guardrails. But even with limits, the onus is on the user to tell others that they are recording, which might not be the case every time.

In defense of this, the argument that might come up is that “You can also record all the time on your phone.” Theoretically, yes. But you would need the phone's mic for other stuff, and if you are trying to be discreet about it, indicators on your phone might give you away.

Apple is rumored to release AirPods with camera sensors next year. That adds another dimension to feed to AI. With the pair of an ever-listening Apple Watch and ever-watching AirPods, Apple thinks it can take on new “context-aware” AI gadgets easily without sneaking in a new form factor.

But there will still be a ton of wearables that would want to compete with Apple’s ecosystem and might make privacy trade-offs to promise more features and personalization. We just don’t know how people feel about these ambient devices being around them, and whether it is even a good thing.

Large Language Models (LLMs) might be the trickiest and most costly part of the voice stack. Companies are now focusing on this part, in addition to developing voice models. Recently, Pipecat’s Kwindala Kramer released PhoneLLM that turns “thinking” mode off for better latency. Last week, Phonely released an LLM trained on over 10 million phone conversations. We might see more LLMs tuned for the voice stack to reduce latency and give more use-case-specific answers.

In Focus

Microsoft and Meta join Google in transcription model wars

In the past few weeks, there has been a barrage of transcription models. Google released its Gemini 3.5 Transcribe a few weeks ago. Wispr released its ASR model, Canto, along with its fundraising announcement last month.

Last week, two big tech companies showed off their new bets in the transcription model kerfuffle: Meta and Microsoft.

Meta released its Muse Voice Transcribe model, which is the company’s first real-time offering. The model was “extensively verified” for over 25 languages and can identify up to 20 speakers, the company claims.

One interesting technological tidbit was that the model spends a bit more time transcribing hard words and speeds up processing for the words it is confident about. We don’t know how that will play out in real-life performance and if it would increase accuracy at all.

On the other hand, Microsoft released its MAI-Transcribe-2 model with a claimed Word Error Rate (WER) of 5.3% over the FLEURS benchmark. It said that the new model processes the audio faster, at a rate of 1 hour of recording in 10 seconds.

Why is there a flurry of new models on the market? There are a few reasons. All companies now have a ton of audio processing to do internally as well as on their consumer-facing surfaces where it offers dictation. For instance, Meta and Google both offer system-wide dictation through their Meta AI and Gemini apps.

Plus, all these companies are integrating more voice features in their coding and productivity apps. All speech recognition models are helpful if you have a ton of meetings and video content to transcribe and analyze.

The biggest market for these companies would be enterprise calling and call centers. Microsoft also dabbles in legal speech capture and powering clinical scribes.

Meta and Wispr are the newer entrants in the market, and the latter doesn’t even offer an API yet. Wispr clearly wants to focus on becoming a transcription layer for hardware and consumer apps. For Meta, the idea would be to power its small and medium business suite for support and consumer outreach.

Benchmarks are interesting from an academic standpoint, but price is what will attract customers. Google is promising a blended price of $0.54 per hour, Meta’s pricing is at $0.18 per hour, and Microsoft is the lowest at $0.1 per hour.

There is an ongoing notion that “voice models are being commoditized” in the voice AI world. But amid this belief, we have seen more models than ever. Low-hanging fruit for developers is driving prices of these models down. On a macro level, companies still want to own model infrastructure even if prices are getting commoditized. Speech recognition is still not perfect in any sense, especially for non-English languages. There is a lot of work to be done.

Quick Bytes

• Google rolls out conversational features across Gmail, Keep, and Docs

Google wants you to talk to its apps. In Gmail, you can search for emails through voice and ask follow-up questions. In Keep, you can refer to notes or add a new note by speaking. Docs is the most interesting one, as its voice assistant can fetch information from Gmail and Drive to create a draft.

• ElevenLabs deploys its AI to Indian appliance company Havells

I am all for voice commands, but this one seems like a tech searching for a problem. Voice commands to turn on/off lights or change the speed of the fan are good, and maybe in India you need regional language support. One feature in this suite is error resolution, which might be better than “agentic commands.” Through this, users can ask what a particular error code shown on the appliance screen means.

• Roland adds generative AI loops, but it doesn’t involve prompting

Music company Roland created a new tool called Melody Flip, which has over 250 gen AI loops based on different music themes. Musicians can import these files into their digital workstations (DAWs) and work on them like editable elements.

• Suno defines its download limits

Suno will start curbing downloads to reduce spam on music streaming services. Free tier users get seven downloads for life, Pro users get 20 downloads per month, and Premier users get 60 downloads per month.

• Bodhan AI launches models for Indic AI

IIT Madras-based lab Bodhan AI launched open-weight models across speech and vision, supporting over 23 languages. This continues the trend of more open-weight models being released to cater to Indian languages.

Numbers Game

• Indian startup Navana AI said it crossed 100 million minutes in automated voice calls

Deals Corner

Wonderful ($550 million): CX funding doesn’t seem to be slowing down as Wonderful hits $5 billion in valuation.

• Investors: Insight Partners (lead), Index Ventures, IVP, Vine Ventures, 9Yards, and Bessemer Venture Partners

Apate.ai ($8.5 million): The company focuses on fraud prevention. It deploys voice and text agents to waste scammers’ time.

• Investors: Lobby Capital (lead), OIF Ventures, Investible, Concept Ventures, and Baobab Ventures.

  ---

Thank you for tuning in. Keep listening.

This newsletter is by Ivan Mehta, a freelance reporter at TechCrunch. It covers AI and technology in voice, audio, and music. Email: voiceaiweek@gmail.com or im@ivanmehta.com

Google is preparing to launch a range of new smart glasses. Apple is rumored to launch its own glasses in the coming years, and OpenAI is working on a smart speaker.

Qualcomm, the chip company, is powering a lot of those new AI devices. Recently, the company’s CEO, Cristiano Amon, said that it is working on over 40 new AI devices with different form factors. In June, it also released platforms for mixed reality and a turnkey toolkit for AI devices, allowing it to white-label its tech.

I sat down with Ziad Asghar, Qualcomm’s SVP for XR, Wearables and Personal AI, to discuss the future of wearables and whether voice is the right interface for it.

What kind of AI devices are you seeing getting traction? Why do you think those form factors are winning over others?

Today, we are seeing amazing pickup on these devices. For example, perhaps the first form factor coming up with AI native devices is smart glasses. With smart glasses, we're seeing amazing traction as they are growing almost 3x every year.

A couple of points to keep in mind. Number one: glasses are an established form factor, as many people wear them daily. I think glasses provide the best input for AI. If you have a camera over here [pointed to the place near your temple], it sees what you see and hears what you hear. I say probably the best form factor for agent AI is a device like glasses.

Second great thing about glasses is that they can actually augment your field of view. For example, when I was in China with one of the customers looking at a demo of glasses based on our Snapdragon AR1, they were speaking in Mandarin, and I could see the full translation coming up on my display, separated by speakers. It shows what is possible with these devices.

How do you think glasses and other devices will work with an AI assistant?

You want a device that’s able to understand the context of your query. Today, you use ChatGPT or some other agent that’s in the cloud, and it doesn’t know why you are asking a question. Devices like glasses can give you context. From my glasses, I can show the assistant a sign or a plaque that I am looking at and augment my prompt with context.

I think our personal AI devices can augment your AI experience. For example, your watch monitors your heart rate and activity level. Glasses can tell you about the weather conditions; your phone knows your location, calendar, and searches. If you can bring and aggregate all this context, you will have the best agentic experiences.

The new AI devices often don’t have screens. How do you think about interacting with them?

We have done research on human-machine interfaces for the last one and a half decades, but from a mixed reality angle. This includes hand tracking and understanding gestures well. We have a good understanding of head positioning and tracking. We have taken these algorithms and hard-coded them in transistors and silicon, which means we can do these use cases at low power consumption. We really have all the pieces to be able to do these human-machine interfaces.

We have worked with a ring maker called Kiwear, which works on both voice and gestures. At times, in public, you don’t want to use your voice for commands.

Voice as an interaction medium is very popular right now. What are some of its limitations?

The first limitation is being able to capture voices in public spaces. Another challenge is filtering out the noise even when you’re not in public.

Voice capture is difficult in noisy environments. How do you handle that?

We are developing some additional unique technologies for this. We have the ability we call the audio bubble. With that, the device only hears sound within a certain radius of the person, even if there is a lot of noise around the user. We have some amazing echo and noise cancellation technologies. So you will not be able to hear anything else except what the person is saying in that small radius.

We are also working on silent speech technologies where nobody can hear what you are saying but the device.

We will need a new interaction mechanism for different environments to work with newer devices. For example, in industrial settings, voice doesn’t work at all.

For smart glasses, there are multiple approaches. There are glasses with cameras that would give AI assistants more context, and there are glasses with LED screens, which can aid productivity. What do you think about that?

What we are seeing is that people are saying the agent that I might need for productivity work is different from the agent I would need for my personal life.

I was on a vacation trip, and I started to take a lot of pictures with my smart glasses. I got used to actually doing that and also started invoking my agent.

[On the enterprise side] We showed this demo for oil and gas companies where you can bring in any equipment, and the glasses can recognize it and give you more information. However, that’s cloud-based AI. We have also shown off demos where small models are running directly on glasses.

I hear once in a while that glasses are gonna replace the smartphone. I really don't think about it that way. You might still have some AI in the smartwatch, in the earbud, in the glasses, and in the phone. And I call this the constellation of devices.

One thing about AI agents and AI applications is they're always context hungry. These devices ask of you that you will provide more context to the agents. But it comes with a privacy trade-off. We've especially seen kind of a restraint towards smart glasses because people don't want to be recorded all the time. With more devices that have the potential to record slyly, some people will cross boundaries of privacy.

I think that companies that have launched glasses with us have been fairly responsible in the sense that an LED light comes on. When you are able to when you take a picture, the LED light blinks, for the most part. When they are making a video, for example.

The second thing is, I see people walking around with the camera like this [made a gesture holding a phone up] in public spaces all the time here in the US now. And there isn't even a light blinking. So my point is, some of that is actually going to get resolved in time.

Qualcomm is very focused on making sure that the consumer is enabled to make that choice [about privacy]. So what we are doing is, for example, in our solution, we call it an always-sensing camera. When my glasses take the picture as the context for AI, it can actually obfuscate all the faces. Because, really, for context, I don't need the faces. What I need is basically all the information that's beyond the faces.

Number three, as we do more of the AI processing on the device, we're actually able to take away the concern of processing data in the cloud.

I asked you earlier about the friction around context capture and privacy. Do you feel that factor is somewhat of a risk to these devices growing? Apart from the protections that you talked about, what is some of the work that you are seeing?

I actually think that the protections will help immensely. But I also do think people are used to other cameras in the public sphere already, like security cameras. And they're capturing people all the time. The same thing goes for phones as well. I feel that, yes, there is a concern, and it's an important thing where we need to do it in a responsible manner. But I think it has not been as big as what I might have expected, which means people are actually getting comfortable with the technology as we provide more of these mechanisms like face obfuscation and other ideas.

Beyond smart glasses, what are some of the form factors that are picking up?

We are working on more than 40 devices of these types beyond glasses. One form factor that I see repeatedly is a pin or a pendant. The reason is that there are people who don't wear glasses. But they still want to have that same sort of multimodality experience with an AI native device.

I also like the idea of headphones with cameras, because a lot of young people wear headphones even in public to listen to music or podcasts. With headphones, there is also space for bigger batteries.

There's a third form factor that I think is very interesting: helmets. In countries like India, there are a lot of people who ride two-wheelers with a helmet. It also has ample room to put a camera or batteries.

Meeting notetakers are still a hot and crucial tool in the productivity stack. Wispr released its own notetaker a few weeks ago, and then scheduling tool Calendly followed suit with its own notetaker. There is no dearth of meeting notetaking tools, yet more companies want to own this part, because transcripts contain action items and hidden prompts that could be used to trigger automation processes, which is the core promise of AI tools.

In Focus

In a survey conducted by caller ID company Truecaller, 76% of respondents in India said that they prefer communicating with businesses through calls. This is a small indicator of why voice AI could be an important business in India, given the potential of automating processes ranging from customer service to appointment booking and sales.

Separately, in a conversation with Pratyush Kumar, co-founder of one of India’s marquee AI startups, Sarvam, VC firm PeakXV’s Rajan Anandan said that voice AI can be a $1 billion category in the country, with calls costing less than ₹5 per minute (5.24 cents), which is below the cost of a human agent in Business Process Outsourcing (BPO) offices.

Depending on the cost of calls, India’s domestic BPO market is estimated to be around $5 billion. The opportunity to serve customers is large, but the race to the bottom on pricing might not be the way. Several voice AI startups in India offer rates lower than ₹5 per minute for automated calls. The challenge with that game is that there is always a bigger fish, which can offer similar capabilities at a lower price.

The argument for voice AI is a bit deeper than cheaper calls. Since AI agents don’t need rest, they can be available at any time. Plus, they take up low-priority tasks that human agents might not get to. Observability platform Last9’s co-founder Kuldeep Dhankar said that agents can take up tasks that humans don’t have time for.

The initial promise of voice AI was handling calls at volume. Then it shifted to providing context-aided customer service. Now, the aim is to automate complex tasks such as user onboarding through voice.

A large market in India could be defined by the crude term “form filling.” Sites and apps in India are hard to navigate. People often find the process of buying something or booking an appointment cumbersome. So they call the helpline numbers and keep waiting until a human agent picks up.

In the same vein, there might be an opportunity to provide government services through voice calls or make it easier for people to access public services. But only a handful of companies will be able to sell their services to governments.

The argument people working on voice, and companies that have adopted voice, make is that there is a need to create new workflows that AI can handle, rather than using existing human workflows, because AI will fail at it.

The bull case is that voice actually becomes an interface, and tasks that took many clicks/taps across many pages can become a simple voice command. This command could be something within the app or could also be a call.

The bear case is that startups in India don’t innovate enough on the technology layer, create differentiation with models, and play the sales game of getting pricing driven down.

Chaitanya Chokkareddy, CTO of CX platform Ozonetel, said that despite startups claiming to offer various capabilities, most customers ask for lower pricing. Plus, enterprise customers are not looking to replace human support and outreach teams just yet. His argument, made in another post, was that a voice agent’s capability to replace a human is not there yet.

One narrative I have heard from voice AI startups is that they are not looking to replace human teams either. They want to deploy voice AI to let agents add to the tally of outbound calls for available leads and incoming support calls.

India is also a tough market because of language barriers. Two people speaking the same language can speak differently. They also mix languages mid-sentence in an unpredictable way. On top of that, users are often in a noisy or unstable network environment.

These are hard technical challenges, and the return on investment might not be that high when you solve for these problems. As Last9’s Dhankar said, voice startups will need to sell at a very high volume to get good margins.

In India, there are 50-60 startups working on voice AI. Some of them are established CX players, and some of them are newer startups working across sectors. For voice startups catering to every industry, there is a chance of losing the price war and burning a lot of cash, or losing contracts to bigger players offering better service packages.

In voice, there are two types of startups that are typically acquisition targets. First, startups that have solved technical problems for voice models, noise cancellation, evals, or similar other areas. Second are companies that focus on a particular vertical, like finance, healthcare, or food, and have had great success in the industry. Current model makers or agent platforms can gobble them up with established sales motion and integrate their products.

India is a potentially large voice market because people like to talk and use voice to connect with businesses and even navigate apps. The country is in the top five revenue-generating markets for voice AI startups like ElevenLabs and Deepgram. If businesses can make it easier to interact with them through voice, there is a big opportunity. But if startups stick to just enterprise calling and price wars, they might be run into the ground.

Subscribe now

Model behaviour

• Daily CEO and Pipecat creator Kwindla Kramer launched PhoneLLM, a fine-tuned NVIDIA Nemotron Nano 30B for voice agents. Kramer’s argument is that often voice stacks use models that use thinking mode to get answers or context, which increases latency. The new PhoneLLM model turns off the thinking.

• Google released Gemini Transcribe 3.5 in both live and batch mode. This model powers dictation on the Gemini Mac app and Rambler tool on Pixel and other Android devices.

  ---

Quick Bytes

• 80 artists in the UK sign a petition for voice rights

Artists including Nicola Coughlan, Hugh Bonneville and Matt Lucas have asked the UK government to take action against unauthorized voice cloning and give rights to their voices. This is an important one to track as more countries across the world are looking at providing protection against unauthorized voice cloning.

• Australia and Finland ban AI music from charts

Australia’s music body said that it is barring fully AI-generated tracks from its official charts. Finland also made a similar move with its new rules. To be clear, these changes are not against using AI, as they still accept songs where there is significant human involvement.

• India’s Gnani launches new models and platform

The Indian AI company launched a new 30B large language model, along with a platform to quickly build voice agents. The model is not specifically tuned for voice use cases, but rather a generic language model.

• Plaud and Boat bet on earphones as a voice AI interface

Plaud launched a new device called Plaud One, which is a pair of earbuds that also have an eSIM in its case to trigger agents. Separately, India’s Boat released a Crest AI platform, powered by the Gemini model, to let users converse with AI assistants. These two are separate roads meeting at one point. One is a meeting note-taking company taking a stab at earphones; another is an audio company taking a stab at AI conversations. It is not clear, though, if having AI capabilities is a pull for consumers.

• Wispr is prepping for an AI assistant and offline support

Wispr is preparing to release an AI assistant that allows users to query all meetings. It also seems to be preparing offline capability.

Call for Sponsors

I am looking for sponsors for this newsletter to support my reporting. I spend hours every week researching different topics and reporting on various slants. I am based out of India, so getting a Stripe account is a very tough task, and that prevents me from turning on subscriptions for the newsletter.

If you want to support independent writing and reporting, get in touch with me at im@ivanmehta.com

Deals Corner

Ringg ($10 million): Indian voice AI company catering to enterprises raised an extension round. The startup also wants to get into other modalities like WhatsApp and chat.

• Investors: PeakXV.

Stability ($76 million): Stability produces a lot of models for image, video, and music generation. But given the investor group, the round seems to be tilting towards sound and music generation.

• Investors: Electronic Arts, Sony Music Group, Universal Music Group, Warner Music Group, and investment firms AMD Ventures and Pacific Alliance Ventures.

Thank you for tuning in. Keep listening.

This newsletter is by Ivan Mehta, a freelance reporter at TechCrunch. It covers AI and technology in voice, audio, and music. Email: voiceaiweek@gmail.com or im@ivanmehta.com

India is a very complex country, and it’s a big market for voice AI because a lot of business is conducted over calls. Yet, in enterprise deployment and model release charts, there are more companies from abroad. That might be changing as Indian companies are beginning to release deployable models.

Last week, Murf AI released the Falcon 2 speech generation model that costs $0.01 per minute. Separately, Blue Machines released the Floe speech detection model, with support for 11 languages and contextual detection. In India, given that there are so many languages, more models with specific language slants and lower cost will find their footing long-term.

In Focus

Why Wispr scored a $2B valuation round

When a certain category of people in tech, namely founders and investors, start using a tool and find it indispensable, the startup is often valued at a high price. We have seen this happen with Granola, the meeting note-taking tool people in tech can’t live without because they fear that they will forget an important thing mentioned in the meeting.

Wispr is in the same boat with its dictation tool, Flow. Voice is the next big interface and is a running thesis within Silicon Valley. While a lot of money in voice AI is concentrated on automating enterprise sales and customer support, meeting note-taking and dictation are two consumer-facing use cases.

The startup’s $280 million round includes tons of new investors such as Acrew, Activate, Forerunner, Goodwater, Peak XV, Together Fund, and PLUS Capital. The company also got money from artists and athletes, which is a big marketing play.

How much a company should be worth is a conversation, a strange combination of math and hope. I want to get into the latter part to understand why investors might think that Wispr is worth $2 billion.

Share

Is good dictation enough?

Dictation itself is not a unique or defensible feature. Wispr Flow has competitors like Willow, Aqua, Superwhisper, Typeless, and Monologue, followed by a swath of indie apps and open-source projects vibe-coded over a weekend. People in tech may move around these tools and might use an open-source project for their own use, but not all products are maintained in the same way, and not all of them have similar features.

Wispr Flow has built a recognizable brand, and a lot of its users are experiencing AI-powered dictation for the first time and, in turn, getting wowed by it. Other tools might have some features a set of users will like, but they don’t have the marketing power of Wispr.

In the past few weeks, many people have complained about Wispr’s transcription quality dwindling. But only a few will completely ditch it to use another tool. The company has accepted that it needs to improve the quality, and as I reported earlier in the month, a new model rollout is on the way.

The company said that the new model, called Canto, will reduce the number of daily edits users have to make to their dictated output by 30-35%. But a speech recognition model unlocks new business avenues for Wispr.

A model for new business models

A few weeks ago, I broke the news of Wispr launching its Granola-like notetaker. Features aside, starting a meeting note-taking tool means that you would need more speech recognition prowess than before. Wispr might still use external models for some speech recognition, but banking on its own model will let it save massive costs.

Powering speech recognition on hardware is another way for Wispr to make money. The company’s CEO Tanay Kothari told me in an earlier conversation that the company is not yet thinking about opening up its API to developers. But the startup is experimenting with this model with a few pilots.

For instance, Wispr Flow partnered with ring maker Oasis to let users directly dictate to their devices like laptops without needing an external mic. Also, one unannounced partnership is Wispr powering speech recognition for Pebble’s Index 01 notetaking ring.

Chasing Jarvis

Wispr started as an idea to build hardware that lets you, well, Whisper. Kothari has talked about his fascination with Iron Man’s intelligent assistant, Jarvis, and chasing a dream to build something similar to that.

The company recently established an Interfaces Lab to build its own models and also explore new interaction modalities with computers. The startup hired Ariya Rastrow, a researcher who was on an early team developing Alexa at Amazon, to head the new lab’s research efforts.

A long-term goal would be to become an all-around assistant while powering speech or voice interfaces in other devices and apps.

Investor speak

Most new investors in Wispr’s round are talking about habit-forming and user needs. Customer-focused fund Forerunner’s Eurie Kim focused on Wispr building tech that reduces keyboard usage.

“When software becomes abundant, what's scarce — what's always been scarce — is understanding the human on the other end well enough that they trust you with something intimate: our thoughts and our voice,” she said.

Meanwhile, Goodwater and Acrew Capital talked about Flow becoming a daily driver, and Together Fund’s Lakshmi Shankar mentioned Wispr having a large swathe of proprietary data for model training.

“Wispr has one of the largest purpose-built datasets of context-labeled audio anywhere, the kind of proprietary, edit-informed data that would cost peers an inordinate amount to recreate,” Shankar said.

Subscribe now

Experiments

More people are building personal tools as voice model makers are allowing generous usage for tinkerers. Voice companies also want to see different use cases that could be deployed commercially.

A person builds a voice agent that calls their friend at 7 am and has a conversation with them to motivate them to go to the gym. This could be great for reminder calls in hotels or appointment booking.

Numbers Game

Wispr’s funding was the biggest story of the week, so let’s dive into some of its numbers provided by the company and investors:

• Users have dictated over 60 billion words with Flow.

• People using Flow cross a 50% voice-to-keyboard usage ratio by month four and are at a 72% voice and 28% keyboard ratio by month five.

• Wispr Flow has retained over 70% of its users after 12 months.

• Roughly 60% of dictation is in non-English languages.

• Over 10,000 enterprises are using Wispr, and users in over 125,000 companies are using the dictation product.

• The company’s revenue has grown by 150% in each of the last four quarters.

Quick Bytes

• Apple finally gets serious about labeling AI music

As the volume of AI-generated songs increases, streaming platforms are thinking more about marking these tracks better. Apple Music, which has adopted a voluntary approach to labeling, will actively start monitoring AI-generated songs later this year, as per Variety. The report said that music labels will need to mark songs with “a material portion of the content” created with AI.

• Meta gets into the dictation game

Wispr Flow got another competitor in Meta in the dictation apps market. The company released a Mac app that can dictate speech in all apps using its own Muse Spark model.

• India’s Sarvam starts deploying in consumer apps

Sarvam, one of India’s biggest AI startups, has claimed to have the best Indic language models for a long time, but its focus has been on enterprise deployment until now. In the past week, the company announced a partnership with HP for a dictation app and also said that it will power call screening on Equal AI.

• Mick Jagger is open to AI-generated music, but not the derivative kind

The Rolling Stones’ Mick Jagger said that he is open to artists using AI as a tool to create music as long as it is original. He said artists should have their own thoughts and composition ideas going into it. “I don’t want people just putting stuff out there that can sound exactly like the Rolling Stones — I think that’s obviously wrong,” he said.

• Adobe adds music, sound, and effects generation to Firefly

Adobe has been working on making Firefly a complete AI-powered creative studio to compete with different companies ranging from ElevenLabs to Canva. Firefly now has the ability to generate music, speech, sounds, and effects within its studio using Adobe’s own and third-party models. Sound and music generation is taking off in marketing; Adobe doesn’t want to lose its place as a creative suite leader.

Deals Corner

Wispr ($280 million): One of the biggest voice deals in recent times to take Wispr beyond transcription

• Investors: Menlo Capital (lead); Acrew, Forerunner, Goodwater, Peak XV, Together Fund, and PLUS Capital (New investors).

Incredible (formerly VISS.AI) ($2.7 million): The startup is developing an agentic assistant that can complete tasks across Mac and Windows

• Investors: Spintop Ventures (lead); Starbright Invest, Magnus Emilson.

HeyBreez ($2.5 million): It’s an enterprise voice agent design and deployment startup with a focus on the Middle East

• Investors: Lunara Partners (lead); Jabbar Group, DASH

Thank you for tuning in. Keep listening.

This newsletter is by Ivan Mehta, a freelance reporter at TechCrunch. It covers AI and technology in voice, audio, and music. Email: voiceaiweek@gmail.com or im@ivanmehta.com

Over the past few years, productivity platforms like Notion and ClickUp have moved into the meeting notetaking space. Now, we are seeing a reverse trend. In the past week, Fireflies announced an email assistant that organizes emails by priority and also drafts replies. Separately, Fathom integrated with Superhuman Go to let users pull up summaries or action items from meetings.

Meeting notetakers already listed drafting follow-up emails as a feature. But with these newer launches, they are moving up the productivity stack. Given how many meeting notetakers and email clients are around, we might see many of them being acquired by productivity companies trying to build a full stack of solutions.

In focus

Deepgram releases a new speech model, and why AI agents need better marketing

Every speech company I speak with says it wants to make its AI voices sound human. Their reasoning is that human-sounding voices can have better conversations with other humans on a call.

Deepgram, which released its Flux text-to-speech model last week, wants to focus on solving customer problems rather than thinking about speech modalities of an agent.

The company said its new speech model, paired with its own Speech-to-text model and context pipeline, can handle user interruptions, changing directions, pauses, and sharing complex information. CEO Scott Stephenson talked about building a robust content pipeline in an interview with Week in Voice AI in May.

“The next phase of this market will not be won by the model that sounds best in a demo. It will be won by systems enterprises can trust to complete business-critical work when conversations become unpredictable,” Scott Stephenson, CEO and Co-Founder of Deepgram, said in a statement.

The company’s new model can respond to a person with just 80ms of latency while retaining context from prior conversations.

While all the model talk is fine, I feel that labs are sometimes exaggerating the need for human tonality in calls to airlines. As a customer, I wouldn’t think, “Oh, that bot sounded so human,” if my query is not resolved. Most customers wouldn’t think about AI’s voices if they get answers. However, many conversations with AI bots are as broken as the ones with human agents. The difference is that people might give other people some grace and expect perfect results from bots.

I am tired of reading marketing material that sounds like “Our AI just works in enterprise, and others suck.” There has to be a better way to convey what your model can do in enterprise settings and for customers. Because if AI was so good, more people would be asking to chat with these bots instead of humans.

Subscribe now

Experiments

I wanted to highlight some voice experiments I spotted in the wild. The core idea in both projects is to solve a personal need and have voice be the layer of interaction for people who are not technically sophisticated. The people who made these projects did the technical heavy lifting and hid the complexities in an efficient way.

1. Dan Peguine built a box for his children to press a button and send a voice message to grandparents on WhatsApp. When grandparents reply to them, the box plays a notification tone, and kids can play the messages at the press of the button. The design is like an old answering machine, which is by design not to involve any screens.

2. Yashaswini Singh made a WhatsApp-based grocery ordering system for her grandmother, who can send voice notes of items she would want to order to a bot. The bot, powered by ElevenLabs voice, will confirm the order and place it. The payment method is cash on delivery, which means that no payment authentication is required.

Both experiments are deeply personal but can resonate with a lot of people while solving for narrow use cases. Voice-only interaction gadgets can reduce screen time for people who are trying to have a digital detox. On the flip side, even if you are interacting with one app (WhatsApp in this case), voice commands can act as prompts to trigger tasks.

The second example is tough to execute at scale given the number of technical layers and APIs involved. If there is no proper error handling, the end user will get confused and/or frustrated. The shopping use case only works when I don’t care about specific prices or brands. When I ask the voice agent to buy rice, it can either pick a brand and quality based on the factors defined on the back end, or it has to know what specific rice I order every time. The moment there is any unreliability, I would switch to ordering directly on the grocery app and ditch the voice agent.

Numbers game

• Deepgram said that it now has $100 million in ARR

• Google said that Gemini now has 1 billion monthly active users, with 63% using voice to chat with the assistant.

• A survey from NextPhone says that AI bots have over 90% resolution rates for callback requests, but have over 55% resolution rate for booking appointments. That means AI agents aren’t yet great at longer conversations with complex turns.

Quick Bytes

• Tyga admits to using AI in his new album

As AI proliferates in music-making tools, we might see more artists using them in the creative process to different degrees. Rapper Tyga, who first said the album was human-made, later admitted that he used AI in the process. It’s not clear if it was AI or just bad music; the album got a 0.0 score on Pitchfork.

• Spotify will now label AI persona on its app

Spotify will label AI personas of artists on the app and will keep the songs away from recommendation algorithms. We might just see the rise of “virtual artists,” just like virtual influencers and a whole lot of business around getting users to stream this music to earn money.

• Wispr Flow is thrown into a privacy debate

A week ago, transcription tool Wispr Flow’s employee posted data about how Indians speak as compared to other countries. This sparked a debate about privacy and user data retention. CEO Tanay Kothari mentioned that the data was based on people who opt in to contributing to model training, and 90% of users opt out during onboarding. It is not clear what data the company retains when they use the data for model training, and whether it removes any PII-related data.

• Google’s Pixel 11 gets useful voice features

Google’s Pixel 11 series gets useful voice features including live translation and transcription for content on the screen. The new phone also has the company’s Wispr Flow competitor Rambler built in.

• Indian payment app PhonePe now lets users ask questions about their expenses

The company launched a conversational voice agent, which lets users ask questions like “How much did I spend on food?” and get visual charts as answers. More apps should let us navigate reports by asking voice-based unstructured queries.

Deals corner

• Wippi ($1.2 million): The Indian company is making gadgets for children that focus on screen-free engagement

  • Investors: 12 Flags Capital (lead), Warmup Ventures, Ventana Ventures, Ruvento Ventures

Thank you for tuning in. Keep listening.

This newsletter is by Ivan Mehta, a freelance reporter at TechCrunch. It covers AI and technology in voice, audio, and music. Email: voiceaiweek@gmail.com or im@ivanmehta.com

Athens-based customer service startup Omilia’s CEO Dimitris Vassos emphasized in a TechCrunch interview that new-age startups like Sierra and Decagon, which have raised millions of dollars in investment, might be deploying too much generative AI and burning through their cash.

His view was that there are a lot of customer service operations that could be automated without deploying AI. Amid the tokenomics debate going on in the tech industry, customer service companies will have to start talking about the cost of automation, rather than the number of calls or tickets they have automated.

In Focus

Boson AI’s quick sprint to its speech-to-speech model, and its avatar-based future

In July, Boson AI announced a big funding round of $70 million, which was originally closed in 2023, to develop new models and deploy its tech to enterprises. Last week, the company released its speech-to-speech model with a claimed latency of 700ms. The model costs $0.0023/min for audio input and $0.014/min for audio output.

The three-year-old startup said that it worked on pre-processing millions of hours of audio data to make the new model efficiently compete with the likes of OpenAI Realtime and Google Gemini Flash.

“Audio itself is fairly cheap. The problem is, this is unannotated audio. We have spent a substantial amount of time to enrich that with the appropriate metadata to give us very high-quality models. Our smaller models are trained on about 10 million hours of data, and for the larger ones, we have about 100 million hours of trainable data available,” CEO Alex Smola, a former AWS scientist, told me over a call.

The company has its own system teams to build a small data center to reduce costs.

“If we were to employ the likes of Scale AI for pre-processing, we couldn't do that as a startup. This is a game that very large, very rich companies can play. As a startup, you have to use math rather than money,” Smola said.

Demo

I got the demo of the company’s newest model, and the conversation bit was very good, and the model seems to handle interruptions well. Though there were a few hiccups, the search tool or API the company was using seemed outdated, as when Smola inquired about Google’s exec team shake-up, the machine wasn’t able to answer that. The second hiccup was when I asked it to pronounce my name as (ee-wan) instead of (eye-wan); it wasn’t able to do that either.

The first problem seems fixable. However, the company would need to fix the second problem for enterprise deployment, as customers can get angry if the system doesn’t pronounce their name in the right way.

Avatars

The company is moving towards avatars and is testing the technology on its API. Smola took a screenshot of me and made a quick avatar, which spoke in an American accent. To me, it was weird and uncanny, but the speech and lip movements were clear. Smola thinks that avatar can be deployed wherever there can be better user engagement.

The company is fighting on many fronts with enterprise deployment and model development. It wants to take on customer service and voice agent companies with its new speech-to-speech models, and it also wants to take on the likes of HeyGen and Synthesia to expand into avatars.

Subscribe now

Experiments

I talked about Vocci’s meeting notetaking ring and in-person meeting notetaking etiquette last week. Days after writing that, I got my Pebble Index 01 ring. The ring is very simple with a microphone and a button. The core idea is to press the button and take notes or set reminders, and I very much like this narrow functionality.

The ring is comfortable to wear, and the gesture of pressing the button to take a note is intuitive. Unlike Vocci, you can send your reminders directly to system defaults like Apple Reminders and organize them later. Given that you can only record while the button is pressed, it is not meant to record meetings.

In my early testing, I was able to record notes in loud environments and also whisper to the ring in a low voice when I was in quiet places like cabs. I think the ring form factor for notetaking might find its own niche.

There are some additional functionalities like double-press to invoke search or an MCP server. One thing I noticed was that Wispr Flow was the ring’s official cloud processing partner. The transcription company seems to be getting into supporting different devices like the Index 01 and the Oasis ring.

Share

Numbers game

• A report by Truecaller suggested that 76% of respondents prefer to talk to businesses over a call. This suggests a big and growing opportunity in India to automate these calls.

• Indiamart, a B2B online marketplace, said in its recent earnings report that it has automated a call volume of 100,000 daily calls.

Quick Bytes

• Suno to adopt watermarking, will restrict downloads to prevent misuse

Gen AI music company Suno’s list of lawsuits is rising, and the company wants to do damage control by adopting measures to let streaming platforms mark songs generated on its platform. The company’s co-founder and CEO Mikey Shulman made it clear that Suno doesn’t want to pass a creative judgement on people using AI to create music.

• EU regulation now mandates disclosure for AI interacting with humans

EU is now asking all companies to disclose their AI system that might interact with humans. This will impact all customer service companies, along with AI avatar companies that work for training and marketing.

• Wispr Flow officially launched its own Granola-style notetaker

As I wrote in my scoop about Wispr’s notetaker launch, the company eventually wants to build a full-fledged voice-operated assistant. Launching a meeting notetaker means that it can gather more context about users, and also help them complete tasks.

• Attackers use voice-phishing techniques to hack top investment firms

Bloomberg reported that hackers used voice-phishing to try and steal information from Point72 Asset Management, Millennium Management, Two Sigma Investments, and Citadel. As more “human-sounding” models hit the market, hackers will try to use them to trick people and companies.

• Josh Talks releases a new India-specific voice benchmark

This is a new benchmark for Indian languages and conditions, which is good for primary model evaluations, but when it comes to deployment, enterprises don’t often look at benchmarks. Mostly because there are so many of them around.

Deals corner

• Omilia ($67 million): The Athens-based CX company raised its second-ever round, as it has reached the scale of 3 billion calls per year. The company is finding a footing in areas like quick-service food chains with clients like Taco Bell

  • Investors: Expedition Growth Capital (lead)

• HappyRobot ($67 million): Another CX funding this week was more focused on infrastructure company HappyRobot, which reached a $1.2 billion valuation. The company has clients like DHL, Kuehne + Nagel, Naturgy, Repsol, and Uber.

  • Investors: Prysm Capital (lead), Eurazeo (co-lead), a16z.

Edit Note (August 17): The story was updated to reflect that Boson raised its round in 2023.

Thank you for tuning in. Keep listening.

This newsletter is by Ivan Mehta, a freelance reporter at TechCrunch. It covers AI and technology in voice, audio, and music. Email: voiceaiweek@gmail.com or im@ivanmehta.com

Despite being embroiled in multiple lawsuits, AI-powered music generators like Suno and Udio have raised millions of dollars in funding.

On the other hand, labels are trying to make the best of the situation and signing deals with platforms, including Suno and Udio, to let them legally use music for remixes and other AI generations to earn money.

Share

Musixmatch, which has been a lyrics provider, is putting its eggs into multiple baskets to be a useful service for search engines, labels, and artists.

I spoke with Rio Caraeff, the company’s co-President of Musixmatch, to discuss how Musixmatch is transforming itself for this age, and how it can do this while supporting artists. Caraeff has a great pedigree in the business as the founder and CEO of Vevo and Executive Vice President, Head of Digital Business at Universal Music.

Caraeff said that over the last year, the company has positioned itself to be a music administration tool rather than just being a lyrics provider. This means that stakeholders like labels can track where their tracks and compositions are being used and how they could be paid from that usage.

To do this, the company has launched a tool called Sentinel, which can keep an eye on copyrighted material use in real-time. This also means that it can track when any material is being used by AI services to generate media like songs. The company announced a deal with Suno for this service on Thursday.

“With Sentinel, our mission is to empower businesses to have total visibility into what’s happening on their platform. If they have an upload or prompt-based generation feature, we can basically determine whether anything is copyrighted or it is in the public domain,” Caraeff said.

He noted that Musixmatch doesn’t want to encourage takedowns of copyrighted content. Instead, he thinks that the opportunity for his company is bigger when platforms keep this content up and use Musixmatch to get a license for that piece of content. He equates Sentinel with YouTube’s Content ID.

“The real opportunity for Sentinel is not taking anything down. The real opportunity is leaving it up. Content ID on YouTube is really the model of how this is a big business opportunity. It’s about leaving things up so you can generate revenue and share revenue with everybody,” he said.

Musixmatch says that the system is used by 250,000 rights owners and can identify even milliseconds of use of material in over 250 languages. On a technical level, it can identify words that are in a prompt or in a recorded waveform.

“This is more robust than just detecting the fingerprint of the waveform itself because the sound can be modified, the sound can be edited, sped up, flipped backwards, and slowed down. And all of those things essentially make fingerprinting brittle there. But the words persist, and they survive,” Caraeff said.

The company wants to evolve Sentinel gradually. In the second stage, it wants to allow the platform to place a request for an instant license, and then, in the last stage, it wants to allow for recognizing full sound recordings.

Musixmatch said that it only passes a confidence score to the platform about the probability of copyright material usage.

Subscribe now

Data deals and artist angles

Beyond Sentinel, the company is also exploring data deals on different fronts. One of them involves paying rights owners through revenue sharing when answers for lyrics and their meanings show up on chatbots like ChatGPT.

What’s more, the company is using its Musixmatch Pro program to let artists opt in, so they get paid when companies train models on their data. Other companies like music creation platform LANDR and distributor Ditto Music also offer similar programs.

When I asked what artists think about competition from generative AI tracks, the former Vevo CEO said that artists worry more about getting distribution.

“The thing that informs our philosophy is that we believe obscurity is the enemy of artistry. An artist’s greatest fear is playing to an empty room. Most artists that we work with want to find distribution. We find that a song being generated that potentially competes with them is a secondary concern,” he said.

Reports paint a contrasting picture, though. A survey of over 4,200 music creators in Australia and New Zealand revealed 82% of them feel AI will prevent them from making a living, but 38% said they are already using AI in their work. There were similar results in the UK, with 66% of respondents saying AI is a threat to their careers.

Some musicians are adopting AI tools while feeling hesitant about what the technology will mean to their careers. Plus, they are also in limbo about when and how much AI they can use without losing audience trust.

“I think that there’ll be an ascendant generation of creators who are AI native. They will have opinions and things to share, and they’re going to express themselves in new ways using these powerful tools. Because they’re just tools at the end of the day. And tools can be used for good and bad,” Caraeff said.

There seems to be a massive demand for AI-generated voices. Companies are trying to sound more “human” than ever, and everyone has a different definition of that. When you try to program “humanness” into AI-generated voices, you can generate a lot of variety, which is important for both creative and enterprise use cases. Investors are ready to back startups in this area as there are a lot of opportunities for both research and market expansion. For companies, there is room to play around with accuracy, expression, and tone of these voices and master one or more categories.

In focus

Wispr Flow is rolling out a new model for quality improvement

Wispr Flow is one of the most popular transcription apps out there. However, the drawback of this popularity is that when something goes even slightly wrong, people start to take notice. Over the last few weeks, many people have complained about Wispr Flow’s accuracy and model quality dropping significantly.

In June, co-founder and CEO Tanay Kothari asked users to express their biggest gripes with the tool. One of the top requests was to improve accuracy.

In a chat with me last week, Kothari said that a year ago, the company decided that it couldn’t just put a band-aid on quality issues with quick fixes. The right way was to introduce a better model that solves current issues.

“We started realizing about a year ago that for these issues, the right answer is not to kind of put a Band-Aid and make those small changes. It comes back to why the model is making these kinds of mistakes, and realizing it hasn’t been trained with more intelligence. To solve this problem, we needed to train a bigger model,” Kothari said.

He said that the new model, which will likely improve quality significantly, is already rolled out to some users, and in the coming weeks will be available to all. The company also plans to better communicate what model changes can and can’t do for users and why there are quality drop-offs for some users and use cases.

New research lab and company direction

Wispr Flow is also taking steps not to be just a transcription company. It hired Ariya Rastrow, a researcher who worked on Alexa, as chief scientist to establish Wispr Advanced Interfaces Labs. Under this project, the company will research new models and new modalities of interaction with AI.

Kothari said that he and his co-founder Sahaj Garg have been thinking about cutting-edge research and new interfaces for a long time. But the challenge was to hire and assemble a team of people capable of working on these problems.

The new lab team comprises 13 people now and will grow bigger. The Wispr CEO said that the company is able to attract top talent as researchers want to work on cutting-edge product problems.

The aim with this lab is to build a new generation of models with reasoning models or intelligence behind them to serve new interaction paradigms.

“I think the whole generations of models that we're going to build are primarily going to be in the human-to-computer interaction space. Right now, it's just voice in, text out, and then you want to add a lot more modalities on both sides. We will use our own architecture and data for that,” Kothari said.

The company is not immediately thinking of building an API business out of its own models. However, it plans to publish more about research findings in the coming months.

As I reported last week, the company is also set to introduce a meeting note-taking feature and take on the likes of Granola, Firlies and Read AI. With this launch, the company could be taking the direction of being an executive assistant.

However, getting into meeting note-taking means the company is set to collect more data from you. Kothari told me that Wispr wants to be respectful and clear about how much context users are ready to give to the app, and if need be, take a step back.

With these moves, Wispr Flow is going to look like a different company, competing in various domains rather than being just a dictation startup.

Experiments

I recently started testing the Vocci ring. My initial impression was that it might be designed for personal note-taking, just like Sandbar’s Stream or Pebble 01 Index ring. However, strangely, it is built more like a meeting notetaker, as you start and stop recording with a button.

The ring is good at capturing audio and transcribing conversations. But the software is not ideal for when you need to take short notes. Second, I like meeting notetakers for in-person conversations to be visible to the other person. With this ring, the recording light is barely visible to me. Plus, it is easy for other attendees not tof notice a ring if a user doesn’t inform them that they are recording the meeting.

While this could count as a feature for some, as we form our meeting etiquette in the AI world, people would be wary of devices like these.

Numbers game

• ElevenLabs reached the mark of $600 million ARR in June. This marks another quarter with the company progressing towards that elusive $1 billion ARR mark. The company’s time to reach the next $100 million milestone has constantly been reducing.

Quick Bytes

• A few weeks ago, I said OpenAI is going to think more about being a strong enterprise player in voice AI. The company released two transcription models in the week called: GPT-Live Transcribe for real-time transcription, and GPT-Transcribe for batch mode.

• RHEI debuted a new agentic platform called Made for Music, which is designed for marketing a new track or an album release. The platform can create the press release, playlist pitch, cover art, and short-form promo assets.

• Friend released a new version of its pendant, which comes with a speaker and a pre-loaded voice that you can’t change. The company told Wired it is using OpenAI’s model to power the companion. We could see more gadgets in similar lines.

• A report by 404 Media highlighted that since Spotify’s AI music labeling is not working as intended, alternative sources such as SoullessMusic.com and Sloptracker have cropped up for users to check for AI-generated tracks. Increasingly, users are demanding more accountability from streaming platforms.

• There is a new class action lawsuit against Granola alleging the app is secretly recording meetings without users’ consent. Granola typically asks users to inform other parties that they are transcribing the meeting. The case highlights the lack of tools that can do this on behalf of users.

• India’s Sarvam released two new models for voice AI: Bulbul v4 for speech generation with better vocal range and Saaras v4 for automatic speech recognition with support for multi-speaker diarization.

Deals corner

Fish Audio ($50 million): The company started as an individual project from an NVIDIA researcher to generate better voices. Largely working in the text-to-speech domain, it has 8 million users and $21 million in ARR

• Investors: Coreline Ventures and Capital Today (lead), 359 Capital, Parable, Play Time, Alphalist Partners, Bayhouse Ventures, Carya Venture Partners, and HF0.

Smallest AI ($13 million): A text-to-speech startup that closed its funding this week. Smallest concentrates on smaller models for specific languages and dialects.

• Investors: Seligman Ventures (lead), Sierra Ventures, and 3one4 Capital.

Encore AI ($30 million): The startup observes voice calls and other customer conversations to train voice agents that can work with call center executives.

• Investors: Team8 (lead), Planven, Lukatz, and Garage.

Thank you for tuning in. Keep listening.

This newsletter is by Ivan Mehta, a freelance reporter at TechCrunch. It covers AI and technology in voice, audio, and music. Email: voiceaiweek@gmail.com or im@ivanmehta.com

The company is set to introduce a meeting note-taker (likely a Granola-style one) in the near future, its changed terms suggest.

The text in the changed terms suggests that the meeting notetaker will transcribe meetings, summarize them, identify speakers, and create notes and action items based on that.

Here is what the changed terms say related to notetaking:

"Notetaker" means Wispr's AI-powered meeting transcription, note-taking, summarization, speaker identification, action item generation, meeting insights, and related meeting assistance features made available through the Services.

The startup is currently testing the tool internally, but there isn’t a fixed timeline for the launch.

There isn’t a dearth of notetakers on the market. There are Granola, Read AI, Fireflies, Fathom, Otter, and Quill. Then there are tons of local, one-time purchase, and open-source notetakers.

With its meeting notetaker, Wispr gets to tell users that instead of paying for transcription and notetaking separately, you can just get one product. Though sometimes, people like an application more if it does one thing well.

Notetaking might be a crucial step for Wispr Flow to be a voice assistant or a full-stack voice company. The startup’s transcription product already learns the way you write emails and messages. With added context from meetings, it can easily create follow-up drafts that are a close resemblance to what you might have written.

Meeting notetaking is a table stakes feature for most productivity apps. It is not hard to plug in a speech-to-text model, transcribe a meeting, and summarize it using LLMs. The North Star for these apps is executives taking action and making decisions based on action items and insights from meeting notetaking. That is a more valuable and sticky asset for startups in the space.

Wispr Flow is valued at $700 million as per the last round. Bloomberg reported that the company is raising a new round at a $2 billion valuation. At that valuation, Wispr would want to be more than just a dictation company.

The startup’s founder, Tanay Kothari, has talked about chasing the vision of Jarvis, Tony Stark’s fictional assistant. That assistant needs to know more about users to execute commands when they talk to it. Context of what users do outside emails and in conversations is important. Meetings are the first step of that ladder to capture additional context for Wispr.

Signal

There are hundreds of hardware and software tools that are ready to “hear” what you say, but they are not yet fully ready to “do” what they heard from you. Voice updates from OpenAI and Anthropic, coupled with Genspark’s meeting notetaker hardware, suggest that companies are busy trying to acquire customers at a voice capture layer. Most of these companies claim users “get work done” through their voice layer, but integration with other software and commands results in final tasks not being smooth yet.

In focus

OpenAI and Anthropic update their voice stack with an agentic slant

Silicon Valley has been chasing the vision of an ever-present and efficient assistant like Jarvis in the Iron Man movies. The assistant listens to you perfectly and executes tasks. We do not have power suit-fetching voice assistants yet, but they can definitely fetch information and reschedule meetings.

Last week, Anthropic updated Claude’s voice mode, which now allows users to choose from Opus, Sonnet, and Haiku. This means that the company updated the intelligence aspect of Claude. Notably, there is no change in the voice model itself, so if you had complaints about how it talks to you, there won’t be any difference.

Another big difference is that the voice assistant can access app integrations like Gmail, Calendar, Notion, and Canva to execute tasks. It can rearrange your meeting or create a design document.

OpenAI, which had updated its conversational chops a few weeks ago, decided to update the ChatGPT desktop app, so the voice assistant can access Codex and Work to trigger agents and have them work in the background without you having to stare at the screen.

The promise of these updates is big, as it allows you to complete work without typing anything or staring at the screen. A startup founder said that he talked to Codex remote on mobile while he was on a hike.

There are a few hindrances, though; you will need to check on agents periodically to see if they are working as intended. For Codex, you will need to remember thread names, as voice runs in a different thread and needs you to call out other agents explicitly.

I partially blame this on OpenAI’s very confusing ChatGPT desktop app, where you don’t know the difference between Codex, Chat, and Work unless you explicitly remember what task you want to complete in what section.

As compared to OpenAI, Anthropic’s voice agent is useful for more general tasks, but there are some teething and latency issues as people switch between models.

AI companies are finally getting their voice features out of the conversation mode, which was useful for brainstorming and saving a transcript. However, with the new features, users can at least begin to do some tasks reliably.

My experiments

I have had brainstorming chats with ChatGPT’s new voice mode on my phone, but I wanted to put it through a different test of longer conversation and memory retention. My gym usually gets crowded during the evening, and I missed my usual time window because of work. I didn’t want to go to the gym, so I asked ChatGPT voice mode to prepare a mixed upper body and cardio workout.

Some exercises were repetitive, and I asked ChatGPT to swap them out. After a bit of back-and-forth, we settled on a routine.

The next task for ChatGPT was to stay with me during the workout and guide me. The first pass I tried was asking it to set an internal timer of 30 seconds per exercise. This went downhill fast. The timer failed multiple times, and ChatGPT told me that 30 seconds were over within 10-15 seconds of starting an exercise block. I pointed out the mistake a few times, and after that, ChatGPT said that it couldn’t do timers.

The alternative I found was that I would run my own timer. Then, once an exercise block was finished, I would ask ChatGPT to tell me the next exercise. This was good during the upper body block. However, after that, ChatGPT moved straight to cooldown rather than cardio, and I had to remind it of that.

The voice mode is good for conversation, but it needs to have a better memory and agentic tools to help with tasks on the phone. I’ll test this once again in a few weeks.

Feature that everyone should use

I recently started re-testing Willow’s dictation app on desktop. One neat feature in the latest update is that when you edit a particular word repeatedly, the app can detect the change and will ask you to add that to the dictionary.

Numbers game

Deezer said that for the first time AI-generated music crossed 50% of daily uploads on the platform. The company noted the volume of songs has gone up to over 90,000 AI tracks every day. Here is how the volume of AI songs has increased on Deezer”

Quick Bytes

• Suno’s troubles continued last week as TechCrunch reported that a November breach contained records of 55.3 million users, according to the Have I Been Pwned website. Separately, thousands of musicians joined a class action lawsuit against Suno and Udio.
  

• Amazon aims to switch models powering Alexa+ conversations as AWS costs could go up to $1.7 billion this year, according to Business Insider. Anthropic models are just not sustainable for Amazon, and it could use its own models.

• LG and Ericsson announced a partnership to bring voice AI directly to the network layer. We have seen Indian telco Reliance Jio make such announcements too. While the tech sounds interesting, we need to see what privacy challenges it can bring.

• Most Meeting notetakers focus on post-conversation knowledge and tips based on the transcript. Now, Otter wants to coach people during the conversation with its Live Assist feature. In the past, we have seen the likes of Cluely use a similar interface for real-time coaching.

• At a $700M valuation, you can’t just be a dictation company. Wispr Flow announced a new Interfaces Lab led by new chief science officer Ariya Rastrow, who worked on Amazon Alexa. He wrote that Wispr will focus on how users can interact with AI better, starting with the voice modality.

Deal corner

• Cast Insights ($4.5 million): A platform that provides audio insights from television, radio, podcasts, and livestreams for newsrooms and consultancies

  • Investors: Abstract Ventures (lead), SRI Ventures (lead), LocalGlobe, Martin Audio, Syndicate (VC)

Partner Spotlight with Atomik Growth

In a chat with a16z, ElevenLabs CEO Mati Staniszewski talks about how he and his co-founder, Piotr Dabkowski, got inspired to solve for badly dubbed movies and created a company that thinks that voice is the new interface for human-computer interaction.

Sponsored content

Thank you for tuning in. Keep listening.

This newsletter is by Ivan Mehta, a freelance reporter at TechCrunch. It covers AI and technology in voice, audio, and music. Email: voiceaiweek@gmail.com or im@ivanmehta.com

Into OpenAI’s hardware play with voice at center

We have seen many rumors about OpenAI’s hardware play with a pair of earbuds and a phone reportedly in the works. Last week, Bloomberg reported that the company’s first device could be a smart portable speaker with a camera that acts like an AI companion.

The speaker will have different sensors to understand your environment, along with a camera. Just like many AI assistants, OpenAI’s idea sounds like a ploy to collect as much context as possible about you and your surroundings in order to personalize its outputs.

More than 900 million people use ChatGPT for work, productivity, travel planning, shopping, and more. OpenAI would hope that the speaker can carry that context over and assist users in all kinds of tasks.

Here is what Bloomberg said about how the device would work:

OpenAI envisions the device anticipating needs, surfacing information proactively and serving as an expert on its user, they said. Though the speaker is designed to stay in the home, it will be easy to move around the house.

The weird bit Bloomberg reported is that the speaker will have mechanical components that move while responding, so users feel that it is alive. That description makes the device sound creepy rather than helpful. I would want my smart speaker to be efficient rather than emotive.

The speaker will directly go up against the likes of Amazon Echo, Sonos, and Google Nest, which have been around for a long time. OpenAI’s bet would be that with more sensors and an assistant that knows more about you, the speaker would attract many users.

The company has already deployed a new voice model, called GPT-Live, which sounds more expressive and human-like. The model is also better at holding conversations with improved turn-taking and interruption handling. The speaker can use a model from that family.

We have heard “AI is the new interface paradigm” many times. For OpenAI, voice might be the modality that unlocks it. The company is working with Jony Ive’s LoveFrom, hoping that the former Apple design head’s experience of creating an iconic new category of devices can wave his magic wand once again.

There are a few hurdles, though. On paper, the ChatGPT app on your phone can do many of the tasks a speaker can do, and there is no clear advantage to having an OpenAI speaker. Most smart speakers don’t have a camera, and in a world where users are wary of AI’s privacy implications, they might not want to have a device that can see what you do all the time from a company that has been accused of privacy infringement.

Smart speakers are good at a limited number of tasks, like setting timers or alarms, adding reminders, or playing songs. They can also control your smart home devices, but that part is unreliable at times. That is because either the speaker doesn’t understand what you said, or there is a breakdown in many pieces of software that connect your bulb to your speaker. AI or no AI, the basic expectation from any smart speaker would be that it can do all of the above without failing.

A more recent hindrance is Apple’s case against OpenAI, alleging that former employees now working at the AI giant stole trade secrets. The court procedure could take a long time, and it might not prevent OpenAI from releasing devices just yet, but there could be future roadblocks, thanks to the case.

AI devices haven’t really taken off in the mainstream yet. OpenAI would hope that it could breach this barrier.

Signals and Experiments

Listeners prefer multi-character AI voices over a human narration: Survey

Spoken conducted a survey of over 1,000 listeners in the U.S. about audiobooks narrated by its AI. The company said that 61% of people mistook AI narration for human narration. Spoken has a Multi-Cast tech with which it can power various characters for book enactments. The company pitted the tech against a single-person human narration.

Key observations from this survey:

• The AI narration had a higher percentage favorability at 61% vs 53% for human narration.

• 31% of listeners said that they would give AI narration a go. The number jumped to 65% after they heard a sample.

• 46% of people who heard the AI version showed an intent to purchase a title, but human narration rated higher, with 49%.

The audio publishing industry reached $2.43 billion in 2025. For smaller publishers or authors who might not have a budget for voice actors, AI could be helpful.

Platforms are offering tools for users to make creation easier. Apple started a digital narration program in 2023. Last year, Audible partnered with publishers by offering over 100 virtual voices to increase its catalog with rapid audiobook creation.

ElevenLabs offers a product for audiobook creation, along with incentives for distribution on its platform. Spotify started accepting books created on ElevenLabs in 2025 and launched a creation platform powered by the voice AI company’s model this year.

Book creation platform Inkfluence AI noted that audiobook creation increased 5x from March to May 2026. As the speed of creation and the variety of voices increase, many authors and publishers might not opt for human narration, impacting voice actor jobs.

Quick Bytes

• Another day, another Suno controversy. 404 Media reported that a hack revealed the company scraped services like YouTube, Deezer, and lyrics site Genius to train its models. Suno previously admitted that it used public sources to get data for its model, but maintains that it is within fair use rights.

• Avatar company Synthesia launched the latest version of its dubbing product, which lets teams refine transcripts and perform edits without spending tokens. The company said that its improved lip-syncing is better at syncing small mouth movements.

• Dictation company Willow made its transcription plan unlimited for base use. The company launched two new models, Willow Frontier Mini and Willow Frontier Pro. The mini model will power the free tier, and the Pro model is for faster and more accurate transcription for paid tiers.

• Local transcription is really useful when your internet is choppy. Communication coaching app Hedy said that it has improved its local dictation abilities on both iOS and macOS.

• Digital audio workstation (DAW) company BandLab acquired an AI music creation studio, Aiode, to expand its licensed audio-to-audio offerings. Aiode claims that its models are created in collaboration with musicians.

Deal Corner

• Rime ($24 million): The voice AI model company raised a Series A

  • Investors: M13 Ventures (lead), Twilio Ventures, Corazon Capital, Unusual Ventures

• Overtone ($18 million): Hinge founder’s new voice-based dating app raised a seed round

  • Investors: FirstMark Capital, Pace Capital, and Match Group.

Partner Spotlight with Atomik Growth

In a chat with a16z, ElevenLabs CEO Mati Staniszewski talks about how he and his co-founder, Piotr Dabkowski, got inspired to solve for badly dubbed movies and created a company that thinks that voice is the new interface for human-computer interaction.

Sponsored content

Thank you for tuning in. Keep listening.

This newsletter is by Ivan Mehta, a freelance reporter at TechCrunch. It covers AI and technology in voice, audio, and music. Email: voiceaiweek@gmail.com or im@ivanmehta.com

Before the deal, the company focused on building emotionally intelligent voice and language models and a pipeline for emotional labeling and metadata from audio. After the deal, under new CEO Andrew Ettinger’s leadership, the company has taken a new direction.

In a chat with me, Ettinger said the company now positions itself as a neutral player by becoming an emotionally intelligent voice infrastructure company.

Hume developed models around semantic space theory, measuring 48 emotions. While the models and pipeline were robust, the company realized that there was high value in its valuation and the data processing framework infrastructure.

“Our infrastructure was something that no one else in the world had because of the level at which we measured and got to that granularity. And so we started licensing that infrastructure out, and have built a fantastic business on top of that in the last five or six months,” Ettinger said.

“You could look at us as Switzerland, where we’re trying to provide infrastructure. The fact that we built this infrastructure with our research for our own models gives us a really unique perspective from first principles to explain and show people how we did this across hundreds of millions of hours of data and human judgments and emotion collections along with an evaluation infrastructure,” he added.

Leaning more into being a model-agnostic infrastructure company, Hume AI is releasing a new benchmark, called Real World VoiceEQ, to evaluate models on emotional intelligence and how “human” they are. The company tested more than 40 models for 60 metrics like Automatic Speech Recognition (ASR), Text-to-Speech (TTS), Speech-to-Speech (S2S), and Speech Understanding. It grounded its results in over a million evaluations, made by 10,000 participants, across the world using its Kairos platform, which is in private preview.

Here are some of the top observations from the study:

• Models can specialize in one ability, like technical accuracy, and falter in another, like emotional intelligence. This means we will see more specialized models. Plus, one-size-fits-all benchmarks are not enough to evaluate different features of a model.

• Voice models have become better at speaking and understanding the emotions of other speakers. However, they still don’t understand nuances like tonality and pacing.

• As we have covered in my story about Treble x Hugging Face benchmarks to measure model performance for real-life scenarios, evaluation metrics often ignore situations like noisy environments, overlapping speakers, and emotional expression.

• One of the biggest challenges for the voice AI industry is speech-to-speech models that listen and speak at the same time.

The study also revealed that even with accent variation in English, models can struggle to understand speech. Here is a passage from the research paper that suggests the error rate goes up 1.5x–2x for South Asian languages:

The accent results show that the native versus non-native gap is the most informative fairness axis. Across selected systems, the native-to-non-native WER gap ranges from +2.15 percentage points for ElevenLabs Scribe to +8.04 percentage points for AssemblyAI Universal-3 Pro. This gap is larger and more diagnostic than a standard-American versus other-regional-accents comparison, because the latter mixes easier native accents with harder non-native speech and hides the underlying fairness issue. Under Kachru’s Three Circles taxonomy, the second-language (Outer Circle) group — Indian, South Asian, Nigerian, Filipino, and related varieties — is the hardest condition for many models. In this subgroup, the top ten systems fall between 6.58% (Scribe) and 12.96% (Modulate Velma-2) WER, while the same systems score between 4.38% and 6.96% on native English.

Hume’s Olya Ossipova said that while studying benchmarks, the company observed that the current way of measuring voice AI models is often just to look at a transcript.

“When you think about expressiveness, what we found is that the model that performs best on expressiveness only ranks fourth in terms of taking action on that expressiveness. So there was actually like a very interesting gap that even if a model understands emotion, it doesn't really know what to do with it,” Ossipova said.

She also mentioned that there is an issue in real-life testing of models where LLM judges use objectivity rather than subjectivity, so you need to perform human evaluations as well. Ossipova said that peering into these problems will give model builders a way to make voice AI models more efficient.

Hume is currently running Kairos, which lets you automate evaluation models for this index, in a private preview. The company is also planning a new partnership with Hugging Face for benchmarks. There is a rising trend of evaluating voice models in different real-life situations, and also assessing models based on their emotional understanding.

OpenAI releases new conversational models to boost voice usage

The headline number from OpenAI this week was that more than 150 million people use ChatGPT’s voice features — voice mode and dictation. Assuming over 900 million people use ChatGPT weekly, that means roughly 16.7% are engaging with its voice features, giving us a clear sense of the massive consumer demand for audio-first tools. This gives us a rough idea of how large the consumer market for voice tools is right now.

OpenAI released a new pair of models last week, named GPT-Live-1 and GPT-Live-1-mini, to enhance ChatGPT’s voice mode for live conversations.

Architectural changes

The company’s naming of the voice model is a bit of a mess because the voice model was previously powered by a model named “Advanced voice model” (AVM). AVM was a pipeline of speech-to-text, a large language model, and text-to-speech. This meant that turn detection and interruption handling were clunky. Plus, you didn’t always get the answers that you wanted.

The company has adopted a new architecture to solve these problems. First, the GPT-Live series is a pair of full-duplex models. This means they can listen and talk at the same time without having to wait for someone’s turn. Second, reasoning, search, and agentic capabilities are separate to keep the conversation going and get access to the best knowledge models like GPT-5.5 for more accurate answers.

App usage

OpenAI has big ambitions with the model. The company said during a media briefing that with this kind of change to voice mode, it thinks of voice “as a kind of primary interface to computing.” With improved conversational skills, OpenAI is betting that its weekly active user base for voice will surge well past 150 million

While OpenAI didn’t provide any data about average session time, the team that built the feature said they had many conversations in voice mode spanning 30-40 minutes. With the new and improved voice mode, the company would like increased session times.

How will OpenAI use these models?

Apple, Google, and Amazon have tried to make their assistants sound more natural, conversational, and knowledgeable. While all these companies have their own devices, OpenAI has to rely on existing platforms for distribution. The new voice mode could set the stage for the company's rumoured devices like earbuds and a phone. Notably, OpenAI was sued by Apple this week over issues of trade theft and breach of contract. It might not derail the AI company’s hardware plans just yet, but we have to keep watching the case to know more.

What’s more, OpenAI plans to bring these models to its API for developers and enterprises. The call center and customer support market is hot. With models that can better handle interruptions, OpenAI could make itself a serious player in the market to compete with the likes of ElevenLabs, Deepgram, and Cartesia.

Signals and Experiments

I got access to the new ChatGPT Voice mode with the rollout and decided to put it to the test. The voice sounds more expressive, and you have the option to switch intelligence (Instant, High, and Medium) and languages directly from a quick settings menu present while using voice mode.

The model still occasionally talks over you, but when it realizes its mistake, it pauses much more naturally than the previous version. Overall, the conversation flowed much better

Earlier in the week, I had a conversation with ChatGPT voice mode about the impact of voice as a modality on app engagement, and the older model wasn’t able to fetch any data. The newer model cited some studies for me.

OpenAI claimed that you can get instant translations while speaking because of the full-duplex nature of the model. In practice, the latency works, but I wish I could say the same about the quality of voice and translation. I tried conversing with the new voice mode in English, and it was speaking Hindi; the pronunciation was that of a foreign-born Indian with a heavy accent, and the assistant spoke bookish Hindi at times that people don’t really use in real life.

The company said that the new voice mode supports “most spoken languages” but didn’t list them. I wish that multilingual coverage were better because I would like to use the assistant for translation when I travel.

The second issue for me was basic work that an assistant can do in voice mode. I asked it to list some points we discussed about newsletter topics, and it started narrating the points without actually listing them. A few times, it told me it could make a framework for graphics for certain stories, but it never ended up making one because it can’t do so.

OpenAI did show off the ability to give visual answers, but those are limited to sports schedules, maps, and weather. I wish the company would give the voice mode the ability to do more things, including creating documents. ChatGPT already has a way to transcribe your prompts, but it would be great if I could keep talking about tasks and the voice mode performs them in the background.

This week, OpenAI launched OpenAI Work, a new agent for project work, and also updated the ChatGPT desktop app to merge Work and Codex. With plugins and coding ability, I would really like to chat with the agent while I might be checking email or doing other tasks.

OpenAI has hinted that it wants people to use their voice for work-related tasks, and the new voice mode could just be the beginning of that path.

Quick Bytes

• Paris-based company Gradium extended its seed round of $70 million in December to another $30 million, bringing the total to $100 million. The company is focusing on building real-time speech models and solutions.

• Music industry bodies have come together to create a program for AI track labeling on streaming services with markers for “AI-assisted” and “AI-generated” tracks. Streaming services like Apple Music and Tidal have adopted their own labeling, while Spotify shows AI involvement in song credits.

• Speechify launched its Simba 3.2 model, which is currently topping the artificial analysis leaderboard. The model also has a competitive base pricing of $10 per million characters, which reduces to $6 per million characters at scale.

• Character.AI is thinking about audio drama with AI characters through its c.ai FM product. The company is currently testing these serialized dramas in a limited way, with professional writers penning stories and AI-generated characters narrating the script.

• Music companies are now trying ethical ways to train models. Music production platform Landr said that more than 30,000 artists have opted into its AI revenue share program. The company has also created a $1 million pool to fund advances for eligible tracks joining the program.

• Google Voice added transcription and note-taking capabilities that already exist within Google Meet. Users can press the “Notes” button during the call to get a summary, transcript, and action items.

Partner Spotlight with Atomik Growth

In a chat with a16z, ElevenLabs CEO Mati Staniszewski talks about how he and his co-founder, Piotr Dabkowski, got inspired to solve for badly dubbed movies and created a company that thinks that voice is the new interface for human-computer interaction.

Sponsored content

Thank you for tuning in. Keep listening.

This newsletter is by Ivan Mehta, a freelance reporter at TechCrunch. It covers AI and technology in voice, audio, and music. Email: voiceaiweek@gmail.com or im@ivanmehta.com

India’s ToneTag bets on its acoustic tech for local processing

After the short era of “Tokenmaxing”, companies using AI are being more cautious about their token budgets and trying to minimize costs. One way that has emerged in this frugal era is to offload part of the processing to local models and chips.

Even if companies can infer some queries without having to go to the cloud, they can start to save on costs. Indian fintech company ToneTag is doing this by processing acoustic signals and answering merchant queries about recent payments locally.

After the Unified Payments Interface (UPI) gained prominence as a digital payment medium in India, banks, payment providers, and fintech startups started deploying small speaker boxes with QR codes. These QR codes “announced” the payments merchants received, which was helpful for sellers who were new to digital payments or might not have a smartphone handy to check the status of a payment.

Using acoustic signals for AI processing

Until now, these boxes have been basic, with some functions like recalling recent payments and volume control. ToneTag, backed by Amazon, MasterCard, and Qualcomm, produces these boxes. The startup recently revealed a new project called eKosha, which uses AI to answer merchant queries.

The startup has a patent for using audio to process payments. At the core, the patent is about encoding and decoding data by embedding it into sound waves. While the original use of this was to facilitate payments between two devices, such as a feature phone and a payment device, the technology could be used to decode requests for language models.

ToneTag’s CEO, Kumar Abhishek, told me that phone processors are capable enough to convert all speech into digital signals like text and hand over the request to models behind Siri or Google Assistant. But for payment boxes, there needs to be a different strategy.

“When you start adding AI to IoT devices, where you have less compute, less power, and less memory, it's very difficult for you to keep listening to requests and process them. We were able to do this because we do analog filtering to make local processing possible,” Abhishek said in an interview.

What can this tech do?

With this implementation, merchants can ask anything about the box itself and troubleshoot problems. Plus, the AI assistant can answer questions about cash flow and recent transactions. This saves merchants from frantically checking bank apps to see the status of a payment or calling customer care to get basic information. The box supports 10 Indic languages and mixed-coded speech with plans for expansion.

ToneTag is also working with banks to put AI to work through cloud queries. With the device, merchants can ask banking-related queries to the device, and also get suggestions for financial products. With all these requests, the bank can gauge who might be interested in lending-related products.

“Because these boxes are conversational, banks could extend their MSME banking right on the box. A merchant can request a checkbook or a credit through voice commands. All tasks that needed a call to the relationship manager or a bank visit could be done directly from the box,” Abhishek said.

ToneTag’s new box has a design quirk for shopkeepers in a dedicated query button. This button serves two purposes: you don’t have to go across the room to start a query, ensuring queries are intentional rather than accidental activations or casual questions about the weather. Given that there is a mandatory button press for the query, merchants have to be close to the device (and, in turn, the mic), resulting in better quality of audio capture. The company said that the merchant's perception of the device and query volume saw an uptick after this implementation.

India has more than 23 million soundboxes deployed across merchants. The problem for fintech companies is merchants retaining soundboxes. Often, distributors offer a box from a company to merchants at a cheaper rate, but as subscription prices increase, merchants would move to another box because the switching cost is low. Abhishek and ToneTag, which aims to reach 500,000 merchants with the new eKosha box in the coming month, think that AI and embedded services will get merchants to stick around with the same box.

Quick Bytes

• Investors are bullish that meeting capture is one of the biggest unlocks of AI. After pouring millions into digital notetakers, they are looking at the physical space where the likes of Plaud and Viaim are competing. Now, another competitor, Pocket, which has sold over 130k devices, got $11 million in funding from Accel, Y Combinator, Peak XV, and ElevenLabs CEO and co-founder Mati Staniszewski.

• ElevenLabs, which has multiple celebrity voice partnerships, will power Gene Wilder’s voice for Netflix’s upcoming reality show “Wonka’s The Golden Ticket.” The company has gotten permission to use the voice from the Gene Wilder Estate, as per Variety.

• Suno is prepping to launch its API for music generation with licensed data. The end goal is to get blessings from the label for partnership and provide other toolmakers with the ability to create sounds or tracks. But legal trouble shows no signs of abating for the company as Winamp Group sued the AI company for unauthorized use of its data for model training.

• ElevenLabs’ value continues to soar. In a fresh tender offer, the voice AI company is looking at a $22 billion valuation, up from $11 billion during its fundraise in February.

• Some reports have suggested that a good chunk of AI-generated music is to game streaming service systems and generate revenue. Tidal is taking a stance against that by demonetizing AI music revenue. The company will tag fully AI-generated tracks, but it didn’t specify what kind of tech it might use to detect AI in songs.

Signals & Experiments

I wanted to escape Delhi’s heat, so I took a two-day trip to Dehradun. While I was there, I stayed in the Taj hotel, and the hospitality was generally great. On the first night, I asked the front desk to send me a toothbrush and a few water bottles. I didn’t get them for 40 minutes. I called them again. This time, I got water bottles, but no toothbrush.

First, I missed my robot friends from hotels in China. Second, can taking requests through voice or chat solve this? I have seen hotels ask you to contact them over WhatsApp through a QR code on the TV screen. But often they just have hotel room information in the chat.

If I can leave a voice note or type my request out, there is less of a chance of missing out on items in my request. This use case doesn’t particularly need large language models. Just a way to get my entry to the request queue to the front desk.

Partner Spotlight with Atomik Growth

In a chat with a16z, ElevenLabs CEO Mati Staniszewski talks about how he and his co-founder, Piotr Dabkowski, got inspired to solve for badly dubbed movies and created a company that thinks that voice is the new interface for human-computer interaction.

Sponsored content

Thank you for tuning in. Keep listening.

This newsletter is by Ivan Mehta, a freelance reporter at TechCrunch. It covers AI and technology in voice, audio, and music. Email: voiceaiweek@gmail.com or im@ivanmehta.com

Why Heidi is building specialized hardware for health

Health-focused voice AI companies have raised over $280 million in funding this year, including Assort Health’s recent $120 million Series C raise to scale voice agents that have handled over 190 million calls. Of that total, nearly $100 million has gone to companies that have a scribe product for notetaking and follow-up actions.

Hardware meeting notetakers are a rising category. Plaud, Anker, Vibe, and Pocket are focusing on professions like doctors and healthcare providers to rely on their gadgets to take notes or record patient sessions. These companies are also targeting all professions where in-person meetings and conversations are taking precedence.

In contrast, Australia-based Heidi Health is taking matters into its own hands by using its specialized models — trained to understand medical terminology better — and approaching clinics directly to deploy its scribe. The company also decided to build its own hardware scribe to improve transcription quality.

I spoke with the company’s CEO, Tom Kelly, about its approach, which has now scaled to handling 3 million weekly visits. Heidi began as a tool for a patient’s history and a chatbot built on top of that.

Building its own models

The company realized that healthcare providers spend a lot of work in admin-related tasks such as follow-ups and filling forms based on information patients provided. That is why Heidi started focusing on making its scribe top-notch. It also wanted to sell directly to clinics to get direct feedback and improve the product.

Kelly said that the company tuned models for medical use and claimed they are better than what frontier labs have to offer for speech-to-text and language summarization.

“In 2025, we struggled a lot with third parties having outages or variations in performance. So we would notice, for example, when a Frontier Lab launched a brand new model, they would be very excited about it. Everyone at Heidi would be disappointed because we would experience a degradation in the old model,” Kelly said.

He added that another problem was that Heidi didn’t have a full view of what happens to requests underneath. That is why the company started building and relying on its own fine-tuned models.

Heidi said that often, generalized models try to “guess” the medical term, which could be dangerous for patients. That is why the startup tries to ground its models in specialized medical terminology for different sectors.

Hardware play

Besides models, Heidi wants to own other parts of its voice stack, including hardware. Rather than relying on just the phone mic or third-party gadgets, the company built its own notetaking device for better transcription quality and, in turn, better follow-up actions.

“The most common cause of bad transcripts is having a bad microphone or a phone that’s put in a bad position. So it’s really important for us to never miss a doctor’s session. We want it to be high quality, and we want to be able to hear what they’re doing,” Kelly reasoned behind building its own device.

Heidi said that healthcare providers often move around a lot, and when the startup tested third-party mics, they weren’t accurate. When it started to test its own device, the results were much better, and it went on to develop its own hardware.

More than just a scribe

The company is also starting to test and deploy its own Wispr Flow-like dictation tool, but for clinical use. The idea is similar: users press a key on their desktop and transcribe what they need across apps.

The company’s aim is to run all of its processing offline in the future. Kelly said that it is possible to run reliable dictation or transcription on a device, but for running LLMs, you either need a dedicated device for a hospital with on-prem deployment or buy a machine with large RAM and storage. It also wants to tap into its origins and reason from a patient’s history to provide a correct diagnosis.

A Menlo Ventures report from 2025 suggested that AI note-taking in healthcare saw a $600 million spend. Investors are very interested in this category as clinics might not switch scribes easily, and startups that build a context graph for health over time will have a dominant market positioning.

Signals & Experiments

Amazon’s Alexa assistant has existed for nearly a decade in India. After its introduction in 2017 with support for English, it later added Hindi. Yes, it understands both languages and sometimes a mix of them. Alexa needs drastic improvements in understanding users.

I have a fan and a table lamp connected to the smart home ecosystem that I can control through Alexa. And often Alexa switches on one instead of another. And at times, it just plays a song instead of switching off the light. I barely ask Alexa for facts or news because it often takes me three to four times to just get my request right.

Last week, I reported for TechCrunch that Amazon is testing Alexa+ in Hindi in India. With this experiment, I hope Amazon is focusing more on accurately understanding Indian speech rather than just adding a “conversational” assistant.

Amazon and Alexa still command a large share of India’s smart home market. However, if you want an assistant to chat with for news, facts, music, and more, there are dozens of options out there with new ones launching every week. Thanks to newer AI models, some of them are good at understanding varied speech. Amazon will need to bring better models to create a stickier experience for consumers.

Quick Bytes

• Coval, a startup built by former Waymo engineer Brooke Hopkins, raised $28 million in a round led by Norwest. Voice AI is scaling up, and enterprises are in need of an infrastructure that tells them how agents are performing and if their money is being used effectively.

• Voice startup Speechify launched its dictation tools for iOS and Mac. With people building their own tools for dictation, the market is heating up. People in the tech industry might choose open-source or niche tools or create one for themselves. However, enterprise customers will tilt towards the likes of Wispr Flow, Willow, and Superwhisper for compliance reasons.

• Zoom is leaning into the ethos of agents working based on meeting notes and discussions with Zoommates. These new agents can create first-draft presentations, reports, project plans, proposals, and spreadsheets from conversations.

• Voice AI company Modulate released a new API to catch AI-generated music. There is a strong wave of AI-generated music making it to streaming services, and companies like Deezer have been fighting it hard with their own tools. Labs like ElevenLabs are adopting Google’s SynthID for generated audio watermarking. But it is still not a widely accepted standard across the board.

Off Topic

I love testing new apps, and I have been obsessed with this iOS weather app called Good Air. It is beautiful and has great transitions as you scroll to different times. Plus, it also tells me the AQI level near me. As someone who lives in one of the most polluted cities. That is important information, and at least I get to look at the horrible weather with a good interface.

Partner Spotlight with Atomik Growth

In a chat with a16z, ElevenLabs CEO Mati Staniszewski talks about how he and his co-founder, Piotr Dabkowski, got inspired to solve for badly dubbed movies and created a company that thinks that voice is the new interface for human-computer interaction.

Sponsored content

Thank you for tuning in. Keep listening.

This newsletter is by Ivan Mehta, a freelance reporter at TechCrunch. It covers AI and technology in voice, audio, and music. Email: voiceaiweek@gmail.com or im@ivanmehta.com

There are independent benchmarks like the Open ASR leaderboard, but they don’t always capture real-world scenarios with messy audio and noise. Plus, these kinds of leaderboards deal with clean audio in a near-field situation where a user is speaking into a phone or a headset.

This is why Hugging Face and Iceland-based voice simulation and testing company Treble decided to create a benchmark specifically for Far Field ASR (FFASR). It evaluates environments where a user is one meter or farther from the mic, such as conference room meetings or someone talking to a smart speaker.

Treble has a simulation where AI companies or device providers can test their models in real-life scenarios, such as various restaurants or cafes. The company told me that over the years, it has developed simulation tech that can create complex situations and test models much faster than before. Using these techniques, Treble can tell AI companies what kind of physics understanding they can build in voice models for better accuracy.

The startup’s CEO, Finnur Pind, said that until now, companies measured their models’ capabilities through the Open ASR leaderboard, which is built on clean audio. He also noted that it could be possible to prepare models for better benchmark performance, as the data is known.

“For Hugging Face and us, it was important to measure how ASR performs out in the real world in complex conditions with overlapping talkers, reverberation, and movement. We wanted to try to create a benchmark that tests these conditions,” Pind said.

The FFASR leaderboard has different datasets for comparison. There near-field speech sample of 8 hours as a base. Then, through Treble’s SDK, there are high, mid, and low signal-to-noise ratio datasets that test models in scenarios like the bathroom, living room, meeting rooms, and hallways with noises like coughing and the HVAC system. Other measurement parameters are of high, mid, and low signal-to-noise ratios with movement. All these tests are based on the Treble10 dataset that the Icelandic startup and Hugging Face collaborated on last year.

However, Pind mentioned that the core data is hidden to avoid gaming of the benchmark. He also mentioned that performance measurement on the old leaderboard had plateaued. And the new FFASR leaderboard will make way for a new kind of measurement.

When model companies test on this benchmark, they would get an average word error rate (WER) across datasets, and an RTFx score, which measures the duration of audio and the time it took for a model to process it. There is also a plot to show the correlation between the two.

“When you test out models on real-world conditions, the performance fluctuation is large for different scenarios. This new benchmark will show companies where their models have been lacking,” Pind said.

Looking ahead, Hugging Face and Treble plan to expand the platform. In a recent webinar, Hugging Face mentioned that in the future, it wants the community to contribute to benchmarking with new open-source datasets for far-field evaluation, ways to test hardware, new denoising algorithms, and different acoustic conditions to test out models. You can check out the benchmark here.

India’s telecom network Jio, owned by the Reliance conglomerate, comes up in every discussion around how it revolutionized data access and made it possible for many users to start using the internet en masse. The company, which is filing for an IPO, is now leaning heavily into voice to make AI accessible to its user base of over 500 million.

The company promises that people using Jio’s network will be able to access AI without needing any extra app and directly on calls. Users will be able to just call up the assistant by saying “Hey Jio” on the call and follow it up with a request to transcribe a call, create action items, place food orders, and more. This raises privacy questions about how Jio is listening in on the call for requests and how much data it retains. Will there be a way to turn off Jio AI for me?

Indian laws would allow users to transcribe a call for personal usage or safety, but usage beyond that without the consent of other people on the call can land you in legal trouble. Also, what happens when you place an international call in a territory where you need two-party consent for recording?

Jio showed a short demo of its voice AI functions during its Annual General Meeting (AGM) of shareholders, and I have questions about how some of that would work in real life.

The demo showed a person in the office on a call, asking Jio’s assistant to add “Rohan, Priya, and Kavya to this call,” and the assistant did its job. But how did it know which Rohan, Priya, and Kavya to add? Did it have access to the person’s phonebook? And if yes, are these the only contacts with these names? It seems unlikely because these are common first names in India.

Moments later, the person on a call with their family asks the Jio assistant to order a pizza. The company said that the assistant knew about the person’s usual outlet without specifying how. It even prompted the person to add Campa Cola, which is a Reliance-owned product. Will the assistant always promote company-owned products? What if I ask it never to recommend Campa Cola?

Jio’s premise is that users in India use voice a lot to chat with each other and with companies. Placing an AI assistant without having them install an app is one of the best ways to get distribution. There are clearly big questions about how the assistant would work, both in terms of execution and data privacy for users.

The company is also introducing voice AI across its apps and ecosystem for mobile recharge, customer support, installation help, roaming and eSIM activation. These use cases are app-based or call-based, so they seem much more plausible and consent-based than the on-network AI assistant.

Jio also showed off a new device called JioTeleFrame, which is a tablet that sits on a wall in your home and shows you reminders, schedules, and weather alerts. Jio calls them agents, but that’s a bit far-fetched. At one point, the demo showed two people talking about planning a festival lunch, and one person asked, “Hey Jio, did you get that?” and the tablet showed a potential lunch menu. Was the tablet listening all along? How would a guest know that a device was listening to them?

The company shows that the tablet can listen to you from a distance, which would be fantastic, because even my Echo devices with multiple mics make mistakes in understanding my voice at a few feet of distance.

The tablet can control the TV, smart home, place shopping orders on JioMart, and possibly talk to other Jio apps and devices. The idea of a central controlling and information device is not new, but in India, where a lot of information is fragmented, it will have to pull a lot of weight to work.

Jio has realized that voice could be a big part of India’s AI story. The company has resources and scale to reach millions of consumers. However, these products will need to perform in real life and not in demos, which would be a tough task to get to.

Signals & Experiments

AI companies are in a race to build AI voices that sound more “human.” These voices would ideally have more texture, timbre, expressiveness, and natural tonality. The core idea is that the more human these AI voices sound, the more likely that actual humans would want to talk to them for use cases ranging from customer support to therapy bots.

It is hard to quantify the “humanness” of a voice model. But that is what Vapi’s new experimental index aims to do. The voice AI pipeline company has built a site where users can listen to a pair of the same voice cloned to different models and pick one based on which one sounds more human. They can carry on doing this for different pairs. Based on these blind votes, Vapi is giving ratings along with a “humanness” score to these models.

At the time of writing this newsletter, xAI’s Grok TTS is the leader, followed by ElevenLabs’ Elevenv3 and Minimax’s Speech 2.5. But these ratings can change as more votes come in. The index also shows latency and pricing for comparison.

Quick Bytes

• There is a new bipartisan bill in the U.S. that would allow citizens to protect their voice from unlawful AI usage. Critically, Google and OpenAI are both supporting this bill. Notably, Denmark gave similar rights to its citizens last year.

• Voice AI company Bland raised $50 million in Series C funding led by Dell Technologies. Fortune noted that the company runs its own models on controlled infrastructure for regulated industries like healthcare. Bland now handles over 3.5 million calls per week and tackled over 175 million calls last year.

• Zoom is adding more voice integration to its suite. Users will be able to use voice commands to start meetings soon. The company also said that this year, it will roll out a feature that will intelligently tag speakers in a call.

• India’s Sarvam, which has multiple voice models and an agent platform, closed a $234 million funding round led by HCLTech. While the company is building large language models from the ground up, it has released multiple voice models that are proficient in Indian languages.

• Hardware voice AI notetaker Plaud said that the company is at $100 million ARR from its software subscription business. It has sold more than 2 million of its note-taking devices. Bloomberg noted that the company is targeting $500 million in sales, and new devices are on their way.

• Apple patched a bug in Beats Studio earbuds that could turn them into a tapping device, according to Malwarebytes. The report noted that attackers could listen in to coversations and also impersonate trusted headphones to other devices.

Partner Spotlight with Atomik Growth

In a chat with a16z, ElevenLabs CEO Mati Staniszewski talks about how he and his co-founder, Piotr Dabkowski, got inspired to solve for badly dubbed movies and created a company that thinks that voice is the new interface for human-computer interaction.

Sponsored content

Thank you for tuning in. Keep listening.

This newsletter is by Ivan Mehta, a freelance reporter at TechCrunch. It covers AI and technology in voice, audio, and music. Email: voiceaiweek@gmail.com or im@ivanmehta.com

Apple’s new Siri AI looks promising, but needs speech improvements

Apple announced the much-anticipated Siri AI (why not just Siri?) at its Worldwide Developer Conference (WWDC). The new Siri is a working iteration of what Apple promised two years ago — an assistant that can better understand you, get insights from the data on the device, look at the screen, and use that context to provide answers.

The company also said that users will be able to customize Siri’s voice by controlling pace and expressiveness. All voice companies and Assistant makers have been trying to get AI to sound more human. The demo of Siri Apple showed in the WWDC keynote video didn’t sound expressive, but this may change as the new betas roll out.

Another small but important announcement from Apple was an improvement in dictation with an on-device model. It’s not clear if the dictation features would equate to using an app like Wispr Flow, Willow, or Monolouge. However, as I have reported before, Apple broke how these apps worked with its iOS 26.4 upgrade. With its own dictation feature built into the default keyboard, users won’t have to switch apps to activate a dictation session.

There is also a dedicated app for the new assistant, which you can use as an alternative to ChatGPT or Gemini. But I haven’t got a chance to test it long enough.

After being on a weeklong waitlist, I got access to the new Siri just today. Here are my very early observations from a few hours of testing:

• Siri can understand me better, even with filler words and complex sentences. However, as soon as I finish one sentence, Siri fires up requests to fetch knowledge even when I am completing my request. When I said, “What are the World Cup matches for today, and where can I watch them?” Siri started searching for matches the moment I said “today.”

• For the above request, for some reason, Siri fetched matches between “Last Friday and Thursday.”

• The voice is not very expressive, but it is better than previous versions of Siri. Right now, the expressiveness dials are in preview, and there is no way to adjust them.

• I turned on the advanced dictation preview, but there is no specific indicator to tell me if I am using the new dictation or the old one.

I am going to keep testing the new Siri for the coming weeks and report my observations.

Model Behavior

Google releases a new live translation model with support for 70 languages

Google dropped a new live translation model powered by Gemini 3.5 with support for over 70 languages. In last week’s newsletter, I talked about using translation apps in China. The new model works on Google Translate, which might not work in countries like China, unless I download the models.

On the enterprise side, the company is debuting the model in Google Meet for enterprises next month. It noted that companies like Grab are experimenting with the new model for customer calls already. This will create a competition between Google and translation providers like Krisp, DeepL, and Palabra

Apple’s new on-device model for powering speech and dictation

Apple’s new Siri is powered by a series of models, including Apple Foundation Model (AFM) 3 and AFM 3 Core Advanced. It’s the second model that is powering the new expressive voice and dictation. The company said that the Core Advanced model is a 20-billion-parameter model, but it acts like a mixture of expert model with 1-4 billion active parameters.

Signals & Experiments

Robot’s slip of tongue

During my China tour, I visited a robotics company called Keenon. They had demoed various robots, and one of them was a humanoid with talking capabilities. When someone asked about a tech blog, it messed up the pronunciation and made it into profanity.

Tech companies might perfect all the robot movements, but numerous examples of chatbots going rogue have taught us that it is highly likely that robots might slip up while conversing with humans at any time. Companies will need to be extra careful as their robots will be out in the wild, and a customer who will experience this mishap might not be as forgiving as tech journalists looking at demos.

Quick Bytes

• Wispr Flow admits that the app’s performance hasn’t been up to date. The company said it is taking steps like scaling its infrastructure to tackle outages and rolling back changes to its autocorrect algorithm. It claimed to have bolstered internal tooling to catch issues quickly.

• India’s Equal AI raises $30 million in funding from Prosus for its AI call screening assistant that supports 10 languages. The app currently handles calls from unknown callers, but it has ambitions to handle calls from known contacts to know the reason. The startup also wants to make outbound calls with its app to book appointments. All of this is very ambitious in India’s multilingual landscape, where not all things are digitized.

• Rylo (formerly Nagish), the platform that helps people with hearing impairments, raised $85 million in funding led by Canaan. The round also has contributions from General Catalyst’s Customer Value Fund (CVF) in the form of growth equity. With the new funding, the company aims to add workplace accessibility, network improvements, and real-time fraud detection to its app.

• Krisp dropped its Voice Translation v3 model with the ability to audit calls live for admins. Plus, the new model has industry-specific vocabulary built in, and it is easier for admins to add custom phrases.

• Deezer has been battling against AI music hard. The company has been offering a tool to other streaming platforms to identify AI-generated music. Last week, the company released a tool for users to identify such music using their playlists imported from other platforms like Spotify and Apple Music. The tool supports detection in 27 languages across 20 platforms.

• Warner Music Group acquired Sureel AI, a tech company that can track how AI is using existing tracks by artists. For labels, AI has presented a new question about how to determine whether the tracks it owns are being used. AI’s usage of data fed for training is fuzzy. There isn’t a standard way to say that a specific percentage of a track was used to generate a new one. But any way for labels to measure usage could be critical for business.

Partner Spotlight with Atomik Growth

In a chat with a16z, ElevenLabs CEO Mati Staniszewski talks about how he and his co-founder, Piotr Dabkowski, got inspired to solve for badly dubbed movies and created a company that thinks that voice is the new interface for human-computer interaction.

Sponsored content

Thank you for tuning in. Keep listening.

This newsletter is by Ivan Mehta, a freelance reporter at TechCrunch. It covers AI and technology in voice, audio, and music. Email: voiceaiweek@gmail.com or im@ivanmehta.com