Meta’s new ASR model tops real-time transcription rankings, costing just $0.18 per hour.

Insight Beating AI Newsflash: Meta Superintelligence Labs has launched Muse Voice Transcribe. The tool integrates real-time speech-to-text, speaker diarization, and end-of-speech detection into a single model. It can process audio up to over one hour long and distinguish between more than 20 speakers simultaneously. In English real-time transcription tests by Artificial Analysis, its Word Error Rate (WER, lower is better) stands at 3.1%, outperforming GPT Live Transcribe (3.9%) and Gemini 3.5 Transcribe Live (4.0%). Muse Voice Transcribe delivers final text roughly 0.16 seconds after a speaker finishes talking. Unlike models that apply a fixed wait time for all words, this model processes 80ms of audio at a time, deciding autonomously whether to continue listening or start outputting. Simple words are transcribed faster, while complex terms wait for additional context to confirm. Meta uses reinforcement learning to optimize both accuracy and latency. The model is trained on over 70 languages, with 25 fully validated, and can directly recognize mixed Chinese-English speech in a single sentence. It is currently integrated into Meta AI for Mac and Muse Code, and the Meta Model API is open for access. Pricing is set at $3 per 1,000 minutes, equivalent to $0.18 per hour.