Meta launches Muse Voice Transcribe, first real-time audio perception model from the super-intelligence lab

Meta announced the availability of Muse Voice Transcribe today through the Meta Model API, Meta AI for Mac and Muse Code. The offering marks Meta’s first foray into real-time speech models and includes streaming ASR, diarization of 20 speakers and more, as well as endpointing, all within a single autoregressive multimodal model from the Muse Spark family, from the super-intelligence lab.

The audio is processed in 80-millisecond chunks (12.5 Hz), each chunk becoming a single soft token. At each step the model decides whether to continue listening using a special token <|next_audio|> or to emit a text token. When the stream ends, a token <|empty_audio|> signals that no further input is forthcoming, prompting the model to complete the remaining text without waiting. This mechanism gives the model full control over how much audio context to accumulate before each word, a property referred to as “delay”.

The central innovation is adaptive delay: the model dynamically adjusts the delay for each word according to difficulty. This is achieved with reinforcement learning (RL) that combines a reward for word-error-rate (WER) and a reward for delay in a multiplicative, not additive, fashion. “The result, according to Meta, is a Pareto front of the speed-accuracy trade-off measured up to final transcription,” the company said.

Building on the streaming ASR backbone, Meta adds special tokens for additional tasks. For diarization, <|start_of_turn|> marks a possible speaker change and <|speaker_{A-Z}|> tags the speaker, with the tag deferred to the end of the chunk. For endpointing, <|speech_onset|> marks speech start and <|speech_endpoint|> marks speech end. Both tasks are trained jointly with ASR and receive dedicated rewards in addition to the base ASR reward.

According to Meta, the model ranks first in Artificial Analysis for streaming speech recognition and public diarization benchmarks as of 1 September 2026. The company has not released its own numerical performance metrics (WER, absolute latency, or results on standards such as LibriSpeech), making the ranking the only available indicator.

The model was trained on 70 languages and up, of which 25 have undergone extensive validation and are recommended as starting points. Native support for code-switching—switching between languages within a sentence or across sentences—is built in rather than added as a separate layer. Context biasing via keywords and language cues is also provided to improve accuracy in specific scenarios.