Meta Launches Muse Voice Transcribe to Beat Rivals

Meta Superintelligence Labs has launched Muse Voice Transcribe, a real-time audio perception model that integrates speech recognition, speaker diarization, and endpointing into a single autoregressive multimodal LLM. On the Artificial Analysis streaming benchmark, the model achieved a 3.1% word error rate. This performance surpasses competing systems, including Cartesia Ink-2 at 3.4%, ElevenLabs' Scribe v2 Realtime at 3.6%, GPT Live Transcribe at 3.9%, and Gemini 3.5 Transcribe Live at 4%.

Unlike traditional setups that chain separate models for transcription, speaker identification, and voice activity detection, Muse Voice Transcribe processes everything jointly. The system compresses 80-millisecond audio chunks into soft tokens at 12.5 Hz, deciding dynamically whether to process more audio or output text. It uses reinforcement learning to manage an adaptive delay, balancing accuracy and latency word-by-word. The model supports more than 70 languages, with 25 validated at launch, handles over 20 speakers simultaneously, and processes hour-long audio files with native code-switching.

For developers and practitioners, this unified architecture eliminates the engineering overhead of managing complex multi-model pipelines. The model is available today through the Meta Model API at a highly competitive price of $3 per 1,000 audio-minutes, which translates to roughly $0.18 per hour. It also powers live dictation features in Meta AI for Mac and Muse Code. On diarization benchmarks, the model achieved a 17.5% average error rate across AMI-IHM, AMI-SDM, and VoxConverse, outperforming rival systems that scored between 21.1% and 28.6%.

Despite these strengths, practitioners should note that the model's weights are not open-source, preventing local deployment or custom fine-tuning. Additionally, the headline 3.1% error rate currently only applies to English speech. However, by treating audio tasks as tokenized text generation, Meta has demonstrated a highly efficient design pattern that could soon expand to other perception modalities.