Meta Superintelligence Labs released Muse Voice Transcribe, the company’s first model for perceiving audio in real time. It transcribes, tells more than 20 speakers apart, and marks the moment a person stops talking, all while the stream is still arriving. The model is live now inside Meta AI for Mac, Muse Code, and Meta’s own Model API, which puts it in front of developers rather than only in a research paper. Meta places it top of the Artificial Analysis streaming transcription table and top of the public speaker-separation boards too, on rankings the company dates to September 1.
The architecture is what makes the real-time claim credible. Meta processes audio in 80 millisecond chunks (12.5 Hz), converts each chunk into a soft token, and lets the model itself decide whether to keep listening or emit a transcribed word. When the model chooses to keep listening it predicts a special token and pulls in the next audio chunk; when the stream ends, a separate token tells it to flush any words still in the queue.
That listen-or-speak decision is where Meta’s speed-accuracy tradeoff lives. Waiting longer before transcribing a word improves accuracy but adds latency, so Meta trained the model with reinforcement learning that combines a word error rate reward with a delay reward, multiplicatively, to let the model vary its wait time word by word based on difficulty. Meta calls this adaptive delay and says it lands the model on the Pareto front of that tradeoff, meaning no faster model in its own testing is more accurate and no more accurate model is faster.
Diarization and endpointing are built on the same token scheme rather than as separate models bolted on afterward. A start-of-turn token marks a probable speaker switch, a speaker tag resolves which of up to 26 voices is talking, and separate onset and endpoint tokens mark when a person starts and stops speaking. Meta trained both tasks jointly with the core transcription task and layered additional rewards on top.
The language claims are the more interesting product bet. Muse Voice Transcribe trained on more than 70 languages, with 25 extensively verified for this release, and Meta says it handles code-switching both within a single sentence and between sentences, plus context biasing so a spoken brand name or contact gets recognized correctly. Diarization past 20 speakers and code-switching mid-sentence are exactly the two failures that make existing transcription tools fall apart in a real meeting or a bilingual household, so the open question is whether the model holds that performance outside the curated demos Meta published alongside the launch.
Meta’s benchmark claims are Meta’s own, measured against Meta’s own model inclusion criteria as of a single date, and the research post does not publish a word error rate figure, a numeric diarization accuracy score at 20-plus speakers, or independent verification from a third party. The post also does not state pricing or usage limits for access through the Model API. Google, OpenAI and several open-weight labs have shipped streaming transcription with speaker labeling before. What is new here is Meta folding diarization and endpointing into one autoregressive model instead of chaining separate systems, and shipping it as a system-level dictation feature rather than a standalone API product first.
Teams building meeting transcription, voice agents, or dictation features should treat Muse Voice Transcribe as a benchmark to test against their own multi-speaker, multilingual audio before assuming Meta’s rankings transfer to their use case.
According to Meta Superintelligence Labs’ research blog post introducing Muse Voice Transcribe, published September 1, 2026.
