Meta is quietly building the next layer of its Muse AI agent, and this time the target is your ears. According to a September 10, 2026 report from TestingCatalog, Meta was testing on-demand custom voices, a Meta AI-style voice mode, hidden reasoning controls, and a new round of “Avocado” model upgrades inside Muse, its personal AI agent. Less than two weeks later, that leak became official: Meta unveiled a customizable “Muse Voice” feature, complete with speed and accent controls, at Meta Connect on September 23, 2026, with Mashable confirming the rollout the following day. Layered on top of a real-time speech model Meta shipped just over three weeks earlier, the combination points to a company racing to close the voice gap with OpenAI, Google, and Microsoft before the end of 2026.
The timing is not incidental. Meta Superintelligence Labs launched Muse Voice Transcribe on September 1, 2026, calling it the first real-time audio perception model built by the lab. Meta followed that, per a Post-Cutoff report, by launching the Muse personal agent itself on September 8, 2026. Nine days after the transcription launch, leaks describing custom voice generation and deeper model upgrades for Muse itself started surfacing, and by September 23-24, 2026, Meta and outlets including Mashable had confirmed the feature publicly as “Muse Voice.” Read together, they sketch a roadmap: ship the plumbing first, then bolt on the personality, then make it official. This piece breaks down what has actually shipped, what only recently moved out of internal testing, and what it means for the AI voice assistant market heading into 2027.
Don't miss new tech stories on Google
Add Tech Insider once in the Google app and our stories appear in your news suggestions.
What Leaked: Custom Voices Coming to Muse
The core of the leak was straightforward. TestingCatalog’s AI Rumours tracker reported on September 10, 2026 that Muse was gaining on-demand custom voices, meaning a user could ask the agent for a specific voice character, such as a robotic voice, and testers found Muse already generating that exact robotic voice and storing it inside a dedicated voice library. TestingCatalog also spotted multiple selectable voices alongside a dedicated voice button inside the Muse interface, according to the same reporting, though public access remained disabled at that stage. Meta had not announced any of this publicly at the time, and the functionality stayed locked to internal testing groups until it resurfaced as an official feature two weeks later.
Alongside the custom voice feature, TestingCatalog flags a Meta AI-style voice mode headed for Muse specifically, plus “reasoning controls” that would let the agent toggle how deeply it thinks through a task before responding. The same September 10, 2026 report identifies two specific internal builds, Avocado 5.14 and a newer Avocado 5.16 v0, running inside Muse test environments, alongside separately tested “MAI” voice models running internally at Meta. None of these internal labels have been confirmed as final product names, but Meta closed much of the gap on September 23, 2026, when it officially introduced the customer-facing feature as “Muse Voice,” complete with speed and accent controls, at Meta Connect. That matters for how this story should be read: what started as a credible leak from a tracker with a strong record on Meta AI rumors has since been validated by Meta’s own roadmap, even if the internal Avocado build numbers themselves remain unconfirmed as official versioning.
Muse Voice Transcribe: The Foundation Already Live
What is confirmed, and already shipping, is Muse Voice Transcribe. Meta Superintelligence Labs introduced the model on September 1, 2026, describing it as a single system that combines streaming speech-to-text, speaker diarization (telling different speakers apart), and endpoint detection (knowing when someone has actually stopped talking) in one pass, according to Meta’s own research blog post. Most transcription pipelines still stitch together separate models for each of those jobs, which adds latency and compounding error. Meta’s pitch is that folding them into a single model cuts both.
Independent coverage of the launch, including a write-up from DataNorth AI, frames Muse Voice Transcribe as multilingual and built specifically for streaming use cases rather than after-the-fact batch transcription. The model carries the identifier muse-voice-transcribe-1.0 and is reachable through a WebSocket realtime endpoint for live audio and a standard multipart upload endpoint for file-based transcription, giving developers two integration paths depending on whether they need live captioning or bulk processing.
Where It Already Runs
On launch day, Meta confirmed Muse Voice Transcribe is already wired into Meta AI for Mac, where it powers voice dictation, and into Muse Code, Meta’s coding assistant, where it handles voice input for programming sessions. That placement is telling. Rather than launching a flashy standalone voice assistant, Meta shipped the underlying transcription layer into existing tools first, treating it as infrastructure other Muse products can build on rather than a consumer-facing headline feature in its own right.
Inside the Numbers: Pricing, Latency, and API Access
Muse Voice Transcribe is priced at $3.00 per 1,000 audio minutes, which works out to roughly $0.18 per hour of processed audio, according to reporting on Meta’s Model API documentation. There is no open-weights release and no self-hosted option on the table. As of mid-September 2026, it is a hosted API service only, which puts it in the same commercial bucket as Whisper API-style offerings rather than open checkpoints developers can run themselves.
On speed, independent coverage describes the model operating on roughly an 80-millisecond cadence for streaming speech, a figure Meta itself has not published as an official benchmark but which trade coverage has cited as the practical responsiveness users should expect in dictation and live-caption scenarios. That kind of cadence is fast enough to feel conversational rather than laggy, which is exactly the bar Meta needs to clear if custom voices and a full voice mode eventually sit on top of this transcription layer.
POST https://api.meta.ai/v1/asr/transcribe
WSS wss://api.meta.ai/v1/asr/realtime
Model: muse-voice-transcribe-1.0
Rate: $3.00 / 1,000 audio minutes (~$0.18/hour)Muse Spark and the Model Upgrade Cycle
The voice layer is only half of what Meta is upgrading. Muse itself runs on the Muse Spark model family, and that family has moved fast in 2026. Muse Spark 1.1 appeared on the Meta Model API listing in July 2026, posting a 43-point improvement over the original Muse Spark on the DeepSWE benchmark, a measure of coding and agentic reasoning performance, according to reporting tied to TestingCatalog’s coverage of the release. Meta paired that upgrade with a public preview of the Meta Model API, opening broader developer access to the improved model rather than keeping it locked to internal Meta products.
Muse Spark 1.3 itself shipped on September 2, 2026, according to a report from Shattered, just days before Meta’s personal AI agent version of Muse launched in early September 2026 running on that same model, as coverage from Yahoo Finance confirmed. Meta’s own blog post on the Spark line describes the model as “a multimodal reasoning model built for agentic tasks, with major gains in tool and computer use, coding, and multimodal understanding,” a description that lines up with the version-to-version cadence: 1.1 in July, 1.3 on September 2, with the custom voice and reasoning-control features now following as the next step in that same upgrade train.
The Ava Model and Hidden Reasoning Controls
Two Codenames, One Unclear Relationship
Separately from the voice leak, TestingCatalog has also reported on a model variant internally referred to as Ava, described in a piece on the Muse superapp as supporting computer use, meaning the model can operate a virtual machine or browser session directly rather than just generating text or audio. A settings toggle for computer use has reportedly appeared in test builds, but Ava itself is not currently accessible to end users. Whether Ava is the same project as the “Avocado” model updates referenced in the voice leak, or a related but distinct effort, is not something Meta has clarified, and no outlet has confirmed a direct link between the two internal codenames.
The reasoning controls mentioned in the same leak would reportedly let users or the system adjust how much computation Muse spends thinking before it acts, similar in spirit to reasoning-effort toggles that have shown up across the AI industry through 2026 as models add slower, more deliberate reasoning modes for harder tasks. Bundled with voice, that combination would let a user ask Muse a quick question in a fast, low-latency voice and reserve a slower, deeper reasoning pass for tasks that genuinely need it, though again, none of this has been confirmed by Meta.
Why Meta Is Racing Toward Voice-First AI
Meta’s Muse is already positioned as an ambitious product independent of voice. It launched as a free, 24/7 personal AI agent running on its own isolated Linux virtual machine with a browser, giving the agent effectively its own computer to complete tasks, according to a September 2026 release roundup. Meta reportedly offers up to roughly 100 million tokens per week free, an unusually generous allotment compared to competing agent products, and connects Muse to Gmail, Google Drive, Calendar, WhatsApp, Instagram, and Facebook Marketplace, along with native iPhone connectors and the ability to spin up subagents for parallel tasks.
Meta has also built commerce guardrails into the agent: a Stripe Link integration issues single-use payment cards so Muse never handles a real credit card number directly, letting it complete purchases like event tickets autonomously while limiting financial exposure. Meta’s coding-agent tier of Muse carries paid plans at $5, $20, and $50 per month, plus a contributor tier where users can opt into having their prompts and completions used for training in exchange for lower usage rates in the $0.10 to $0.20 per million token range. Layering a natural, customizable voice on top of an agent that already handles email, messaging, shopping, and payments is the obvious next step if Meta wants Muse to feel less like a chat window and more like an assistant you talk to the way you would a person.
Market Impact: What This Means for Meta’s AI Business
Smart Glasses as the Real Target
Meta’s voice push is not confined to phones and laptops. Reporting on Muse Spark’s rollout describes voice features, including faster voice replies, live camera-based recognition, and a shopping mode, already reaching Ray-Ban Meta and Oakley Meta glasses through 2026. On September 23, 2026, Meta announced that Muse would come to its AI glasses as a personal agent, with hands-free access and the ability to act on what the wearer is viewing, and on October 1, 2026 it said Muse was “coming to” its AI glasses in the coming months and could be activated by saying the agent’s name. That is the strategic bet underneath all of this: Meta owns a hardware surface, smart glasses, where a screen-based chat interface barely works but a voice interface is close to mandatory. If custom voices and a fuller voice mode make it out of internal testing, glasses are the most likely place Meta pushes them first and hardest, since typing or reading text on a lens is a much worse experience than simply talking to the device.
There is also a straightforward developer-monetization angle. By pricing Muse Voice Transcribe as a metered API product at $3.00 per 1,000 minutes rather than folding it silently into a subscription, Meta is treating voice infrastructure as a revenue line the way OpenAI and Google already do with their own speech APIs. Every product Meta builds on top of that transcription layer, custom voices included, inherits the same usage-based economics, which gives Meta a second monetization path beyond advertising and subscription tiers if the agent business scales.
Competitive Landscape: Muse vs. ChatGPT Voice vs. Gemini Live
Meta is not entering an empty market. OpenAI’s ChatGPT has offered a conversational voice mode across its apps for well over a year, and Google’s Gemini Live provides a similar real-time, conversational voice and video experience baked into the Gemini app on Windows and Pixel devices. Microsoft has pushed its own voice experiences through Copilot on Windows and in Microsoft 365. None of these competitors has publicly detailed a user-facing custom voice cloning feature comparable to what is leaking out of Muse, which is where Meta’s approach, if it ships, would differentiate itself: rather than offering a fixed set of assistant personas, Muse’s leaked design lets a user generate a specific voice on demand and save it to a personal library.
Meta’s stated technical edge so far is on the transcription side rather than the generative voice side. The 80-millisecond streaming cadence attributed to Muse Voice Transcribe, combined with combining speech-to-text, diarization, and endpoint detection into a single model, is a genuine architectural difference from pipelines that chain several specialized models together. Whether that translates into a better end-user voice assistant experience depends entirely on the generative voice layer Meta has not yet shipped, which is exactly the part still sitting in internal testing.
Historical Context: How Meta Got Here
Meta’s voice ambitions did not start with Muse. Meta AI, the assistant embedded across WhatsApp, Instagram, Facebook, and Messenger, had already rolled out faster voice replies and live camera-based recognition earlier in 2026 as part of the Muse Spark rollout, according to TestingCatalog’s coverage of that release. That earlier work effectively built the audience and the app-level plumbing; what changed in September 2026 is that Meta gave the underlying speech technology its own name, its own model number, and its own API pricing, signaling that voice has graduated from a feature bolted onto existing apps to a standalone product line Meta expects developers to build on.
The broader industry backdrop matters too. Reasoning-effort controls, the kind referenced in the Muse leak, became common across major AI labs through 2025 and into 2026 as models added slower, deliberate thinking modes for harder problems. Meta folding a similar concept into a voice-first agent suggests the company is trying to import lessons from the text and reasoning side of AI development into a voice interface, rather than treating voice as a separate, simpler product track.
What Meta and Industry Voices Are Saying
Meta has described Muse’s mission in broad terms rather than detailing the voice roadmap specifically. In its official announcement, Meta called Muse “a secure, private personal AI agent that proactively helps people meet their goals and suggests ideas”. Mark Zuckerberg, Meta’s chief executive, went further in comments cited by The New York Times, describing Muse as working “24/7 on your behalf to help achieve your goals and improve your life, your health, your relationships, your finances.”
A Meta executive spoke to the agent’s extensibility in comments reported by Yahoo Finance, explaining that anyone with a service they like that has an API can ask Muse to build a custom connector for it. The same Yahoo Finance report described Muse as powered by Muse Spark 1.3 and able to tap into apps such as Instagram and WhatsApp. Meta has since announced Muse Wonder and Muse Wonder Full on October 12, 2026, and said Muse Wonder would soon power Muse, initially for subscribers.
On the model side, Meta’s own engineering blog described the Spark line’s design goals directly, stating that “Muse Spark 1.1 is a multimodal reasoning model built for agentic tasks, with major gains in tool and computer use, coding, and multimodal understanding.” That statement does not address voices, but Meta’s Connect materials on September 24, 2026 said users could design a Muse voice by describing preferences such as a faster or slower delivery or a particular accent.
Meta Muse Voice and Model Timeline
| Date | Milestone | Status | Source |
|---|---|---|---|
| May 2026 | Muse Spark voice mode and camera recognition roll out to Meta AI, Ray-Ban Meta and Oakley Meta glasses | Shipped | TestingCatalog |
| July 2026 | Muse Spark 1.1 listed on Meta Model API, 43-point DeepSWE improvement | Shipped | TestingCatalog / Meta |
| September 1, 2026 | Muse Voice Transcribe launches as Meta’s first real-time audio perception model | Shipped | Meta Superintelligence Labs |
| September 8, 2026 | Muse personal agent and iPhone app launch, powered by Muse Spark 1.3 | Shipped | 9to5Mac / Yahoo Finance |
| September 10, 2026 | Custom voices, reasoning controls, and Avocado model updates reported in internal testing | Leaked, unconfirmed | TestingCatalog |
| Not disclosed | Public rollout of Muse voice library and voice selector | No confirmed ship date | TestingCatalog |
Voice AI Comparison: Muse vs. the Field
| Product | Company | Voice Capability | Distribution | Notable Detail |
|---|---|---|---|---|
| Muse Voice Transcribe | Meta | Real-time speech-to-text, diarization, endpoint detection | Meta Model API, Meta AI for Mac, Muse Code | ~80ms streaming cadence, $3.00 per 1,000 audio minutes |
| Muse custom voices (leaked) | Meta | On-demand voice generation, saved to a personal voice library | Internal testing only | Disabled for public users, no confirmed release date |
| ChatGPT Voice Mode | OpenAI | Conversational voice chat | ChatGPT mobile and desktop apps | Bundled into existing ChatGPT subscription tiers |
| Gemini Live | Real-time conversational voice and video | Gemini app, Android, Pixel devices | Bundled into Gemini subscription plans | |
| Copilot Voice | Microsoft | Voice-driven assistant chat | Windows, Copilot app, Microsoft 365 | Integrated into the Microsoft 365 Copilot experience |
What Still Isn’t Confirmed
It is worth being precise about what this story is and is not. Muse Voice Transcribe, Muse Spark 1.1, Muse Spark 1.3 (which shipped September 2, 2026), the Muse personal agent, and the Muse coding agent’s paid subscription plans (an October 14, 2026 report lists four) are all real, shipped, and documented by Meta or by outlets covering Meta’s own announcements. Custom voices are a different case. No release of the feature has been confirmed as of October 22, 2026. Readers should treat custom voices as an unconfirmed direction, not a guaranteed product.
Predictions: Where Muse’s Voice Layer Goes Next
Based on the pattern Meta has followed with every other piece of the Muse stack, a few outcomes look likely over the next two to three quarters.
- Meta will likely ship a limited or regional preview of custom voices before a full global rollout, mirroring how Muse Voice Transcribe first appeared inside Meta AI for Mac and Muse Code rather than as a standalone consumer launch.
- Smart glasses, specifically Ray-Ban Meta and Oakley Meta, are the most probable first hardware destination for any finished voice mode, since Meta has already been routing Muse Spark’s voice features to glasses throughout 2026.
- Expect Meta to keep metering voice features through the Meta Model API rather than bundling them free into Muse, continuing the usage-based pricing pattern set by Muse Voice Transcribe’s $3.00 per 1,000 minute rate.
- The Ava computer-use model and the leaked reasoning controls will likely converge with the voice work, since an agent that can both operate a computer and hold a natural voice conversation is the clearer end goal than either capability alone.
- Competitive pressure from OpenAI and Google will likely accelerate, not slow, Meta’s voice timeline, given how directly the leaked features target the conversational voice territory ChatGPT and Gemini Live already occupy. That same pressure is already visible in how fast rivals like xAI’s Grok have iterated their own agentic and coding models through 2026.
None of these are guarantees. They are reasonable extrapolations from a company that has, in the space of about four months in 2026, gone from voice replies inside WhatsApp to a named, priced, API-accessible speech model, with custom voice generation now the next visible step in that chain.
Frequently Asked Questions
Is Meta Muse getting custom voices right now?
Yes. After TestingCatalog’s September 10, 2026 report described custom voice generation in internal testing only, Meta officially introduced the feature as “Muse Voice,” with speed and accent controls, at Meta Connect on September 23, 2026, a launch Mashable confirmed on September 24, 2026.
What is Muse Voice Transcribe?
Muse Voice Transcribe is a real-time audio perception model Meta Superintelligence Labs launched on September 1, 2026. It combines streaming speech-to-text, speaker diarization, and endpoint detection in a single model, and is available through the Meta Model API, Meta AI for Mac, and Muse Code.
How much does Muse Voice Transcribe cost?
Reported pricing is $3.00 per 1,000 audio minutes, or roughly $0.18 per hour of processed audio. There is no open-weights or self-hosted version; it is a hosted API service only.
What model powers Muse?
Muse’s personal agent launched in September 2026 running on Muse Spark 1.3, according to Yahoo Finance’s coverage of the launch. Muse Spark 1.1, an earlier version, shipped in July 2026 with a 43-point improvement over the original Muse Spark on the DeepSWE benchmark. Meta announced Muse Wonder on October 12, 2026 and said it would soon power Muse, initially for subscribers.
What is the Ava model Meta is testing?
Ava is an internally tested model variant described by TestingCatalog as supporting computer use, meaning it can operate a virtual machine or browser session. It is not currently accessible to end users, and Meta has not confirmed whether it is directly connected to the leaked “Avocado” model updates.
How does Muse compare to ChatGPT Voice and Gemini Live?
ChatGPT Voice Mode and Gemini Live both offer conversational voice chat bundled into their respective subscription apps. Muse’s leaked design differs by letting users generate custom, on-demand voices saved to a personal library, a feature neither OpenAI nor Google has publicly detailed for their own assistants.
Will Muse’s voice features come to Ray-Ban Meta glasses?
Meta has already routed earlier Muse Spark voice features, including faster voice replies and live camera recognition, to Ray-Ban Meta and Oakley Meta glasses during 2026. Glasses are considered a likely early destination for any finished custom voice mode, though Meta has not confirmed this specifically.
Is Muse free to use?
Meta’s personal AI agent version of Muse is free to start, reported at up to approximately 100 million tokens per week, with paid Power ($20 a month for 500 million tokens a week) and Maximum ($100 a month for 3 billion tokens a week) plans. It is available in the United States and Canada. A separate Muse coding agent has its own paid subscription plans, and an October 14, 2026 report found no free plan, trial, or free credits for it, alongside a contributor tier with reduced per-token rates for users who opt into having their data used for training.
