Meta launches Muse Voice Transcribe for live speech-to-text

Meta Superintelligence Labs unveiled Muse Voice Transcribe on September 1, 2026, its first real-time audio perception model. The system transcribes speech, distinguishes more than 20 different speakers within the same recording, and detects the exact moment a person stops talking — three tasks that until now required three separate systems, one for transcription, one for diarization and a third endpoint detector, stitched together with the latency and failure points every hand-off adds. The model is already live through Meta's API as muse-voice-transcribe-1.0, and it now powers voice dictation in the Meta AI app for Mac and in Muse Code, the company's AI coding assistant. Meta has not released the model's weights, so it can only be used as a hosted service, with no option to run it on your own servers.

How it works: listening and writing in one loop

Muse Voice Transcribe belongs to the Muse Spark family and processes audio in 80-millisecond chunks, or 12.5 chunks per second. Each chunk becomes a token that the model analyzes autoregressively, inside the same decoder loop that also generates text — there is no separate alignment stage between what is heard and what is written, which is exactly where traditional pipelines tend to drift out of sync. After every chunk, the model makes a binary choice: it either predicts a token asking to keep listening, or it emits a word directly. When the audio stream ends, a special token tells the model no more input is coming, and it immediately flushes any pending text — a behavior Meta calls 'flexible listening,' which it also applies to decide who is speaking at any given moment.

A delay trained, not fixed in advance

How much audio the model waits for before writing a word — what Meta calls 'delay' — is not a fixed value: more context makes the transcript more accurate, but it also adds latency. Instead of hand-tuning that trade-off, Meta's engineers trained it with reinforcement learning, multiplying a word-error-rate reward with a speed reward so the model itself learns to vary the delay word by word, based on difficulty. To identify speakers, the system does not rely on a separate model but on special tokens inside the same stream: one marks a possible turn switch as soon as it happens, and another, slightly delayed, assigns a speaker tag from A to Z; audio from one speaker can be split across several segments that all resolve to the same tag. Two more tokens mark the start and end of an utterance for the endpoint detector, and all three tasks are trained together on top of the same base transcription reward.

Results ahead of Cartesia, ElevenLabs and Gemini

• 3.1% word-error rate on the final transcript, 0.16 seconds after the end of speech, on the Artificial Analysis AA-WER Streaming benchmark — ahead of Cartesia Ink-2 (3.4% at 0.43s) and ElevenLabs Scribe v2 Realtime (3.6% at 0.14s).
• 3.6% error on the first partial transcript, at just 0.13 seconds; Cartesia Ink-2 with external endpoints is faster (0.07s) but less accurate (4.0%).
• 17.5% average diarization error rate across the AMI-IHM, AMI-SDM and VoxConverse benchmarks, versus a range of 21.1% to 28.6% for the five other systems tested.
• Supports recordings longer than one hour and more than 20 simultaneous speakers with no post-processing required; accuracy can be further boosted with language, keyword and context biasing.
• Priced at $3.00 per 1,000 minutes of audio processed ($0.18 per hour), undercutting Cartesia Ink-2's $4.00 and less than half of ElevenLabs Scribe v2 Realtime and Deepgram Flux's $6.50.

The model was trained on more than 70 languages, 25 of which were extensively verified before launch, and it supports switching languages mid-sentence without losing accuracy, something common among bilingual speakers. Meta announced the release on September 1, 2026 through its @AIatMeta account on X, describing it as its first real-time perception model — arriving before OpenAI or Google have shown an equivalent live-transcription system of their own. Anyone with the Meta AI app installed on a Mac can already try it: just hold down the Fn key and speak in any application to have the dictation transcribed with the new model.

For companies and professionals in the United States, the United Kingdom and other English-speaking markets who record meetings, press briefings or interviews with several people talking at once — journalists, law firms, customer-support teams, accessibility services for deaf users — a model that transcribes, separates voices and detects pauses in a single pass cuts infrastructure costs and the error margin of chaining several tools together. In exchange, it requires accepting that the model can only be used through Meta's API, with no option to self-host it or audit its weights. Before replacing a transcription pipeline already in production, it is worth checking the real error rate against each team's language and accent, since only 25 of the more than 70 supported languages have gone through extensive validation.