Meta Superintelligence Labs launched Muse Voice Transcribe on Tuesday, the lab's first real-time audio perception model, combining streaming transcription, speaker diarization for more than 20 speakers, and endpoint detection in a single pipeline priced at $3.00 per 1,000 audio minutes ($0.18 per hour)1,2.
On the Artificial Analysis AA-WER Streaming speech-to-text accuracy benchmark, Muse Voice Transcribe recorded a 3.1% word error rate, ahead of Cartesia Ink-2 at 3.4%, ElevenLabs Scribe v2 Real-time at 3.6%, GPT Live Transcribe at 3.9%, and Gemini 3.5 Transcribe Live at 4.0%. The benchmark covers English speech only. Meta said the model took the number-one spot on the Artificial Analysis streaming evaluation as of September 1.
The model was trained on more than 70 languages, with 25 extensively validated for the initial release4. It supports multilingual code-switching mid-conversation and long audio sessions exceeding one hour. Meta also said the model supports language and keyword biasing, and that ASR, diarization, and endpointing are trained together rather than running speaker clustering as a separate downstream process.
Muse Voice Transcribe is an autoregressive multimodal model from the Muse Spark family. Audio arrives in 80-millisecond chunks, each compressed into a single soft token; at every chunk, the model either emits a text token or a special placeholder indicating it needs more audio context. Meta calls this mechanism "adaptive delay," where difficult words receive more context while easy words are transcribed almost immediately. The accuracy-latency tradeoff is learned during a reinforcement learning phase in which word-error-rate and delay rewards are multiplied rather than added.
The 20-plus-speaker diarization ceiling places Muse toward the high end of the market, though it is not the highest published figure. Speechmatics' real-time transcription service documents identification of 50 speakers by default and up to 100 with configuration changes, while Amazon Transcribe specifies a maximum of 30 unique speakers for streaming, according to a vendor documentation review by VentureBeat. Meta reported an average 17.5% diarization error rate across the AMI-IHM, AMI-SDM, and VoxConverse benchmarks.
Muse Voice Transcribe is available via the Meta Model API, Meta AI for Mac, and in Muse Code. Streaming and non-streaming transcription cost the same, and Meta said zero-data-retention processing is priced at parity with standard processing. The API currently provides eight concurrent streams per tenant by default, with real-time sessions of up to 60 minutes before an application must reconnect.
Unlike Meta's Muse Glimmer models, the company will not release open weights for Muse Voice Transcribe, a Meta spokesperson told The New Stack.
ANALYSIS The closed-weight decision marks a departure from Meta's usual open-release posture and limits third-party fine-tuning or self-hosted deployment. Pricing at $0.18 per hour, combined with integrated diarization that removes the need for a separate speaker-clustering step, compresses what has traditionally been a multi-vendor pipeline into a single API call.