Production voice stacks usually run three stitched systems: one model transcribes, a second separates speakers, a detector decides when the user stopped talking. Every hand-off adds latency and a new failure mode. On September 1, Meta Superintelligence Labs released Muse Voice Transcribe, collapsing those three jobs into one autoregressive model — what the lab calls its first real-time audio perception model.
One Model, Three Jobs
Technically, Muse Voice Transcribe comes from the Muse Spark family. Audio is processed in 80ms chunks (12.5 Hz), each transformed into a single soft token; at each chunk the model decides whether to keep listening (<|next_audio|>) or emit a text token. Diarization and endpointing are not bolted-on modules either: extra special tokens (<|start_of_turn|>, <|speaker_A|>, <|speech_endpoint|>) carry them inside the same sequence. Meta says all three tasks are trained jointly, with no required post-processing.
The most interesting design choice is "adaptive delay": waiting longer before transcribing improves accuracy but raises latency, so Meta used reinforcement learning to combine a word-error-rate reward and a delay reward multiplicatively, letting the model decide per word how much audio context to wait for.
The Numbers
Per Meta's published charts, final-transcription streaming word error rate is 3.1%, with seven other systems in the same chart ranging from 3.4% to 4.0%; average diarization error rate across AMI-IHM, AMI-SDM and VoxConverse is 17.5%, versus 21.1%–28.6% for five other systems. On the speed-accuracy scatter, the model reaches about 3.0% error at roughly 0.16 seconds, below the previous Pareto frontier formed by Soniox, Cartesia and ElevenLabs systems. Meta also states the model ranks first on Artificial Analysis streaming speech-to-text and on public diarization benchmarks as of September 1, 2026 — a vendor claim, with the numbers traceable to the charts above.
Independent coverage from MarkTechPost adds context: on the same board, Cartesia Ink-2 (semantic endpoints) sits at 3.4% WER / 0.43s and ElevenLabs Scribe v2 Realtime at 3.6% / 0.14s; pricing works out to $0.18 per audio hour, below Cartesia's $0.24 and ElevenLabs' $0.39, with an API-only release and no open weights so far (these three items are single-sourced to MarkTechPost).
Languages and Availability
The model is trained on 70+ languages, 25 of them extensively verified and recommended for the initial release. It natively supports arbitrary code-switching within and between sentences — one demo is a long Mandarin-English monologue about running Muse Glimmer locally on an RTX 3090 with 4-bit quantization that transcribes cleanly. Long audio over one hour with 20+ speakers is natively supported; the official blog posts a full transcript of a 1-hour-52-second, 11-person real conversation. Availability is via Meta Model API, the Meta AI for Mac client, and Muse Code, where holding the Fn key gives system-wide voice dictation.
Two Things Worth Noting
First, this is an API-only release with no open weights — a contrast with Meta's earlier generosity in open-sourcing Muse Glimmer 30B and the Muse Spark line. The "personal superintelligence" story starts with closed ears. Second, folding ASR, diarization and endpointing into one token space mirrors what text models did with tool calls and reasoning traces: shorter pipelines, better real-time behavior, fewer failure modes. Teams building voice agents should read the official blog end to end (source: https://research.meta.ai/blog/introducing-muse-voice-transcribe).
Once transcription, speaker separation and turn-taking are just special tokens inside one model, the next bottleneck for voice interaction may no longer be whether the machine can hear clearly.