Meta Launches Muse Voice Transcribe for Real-Time, Multi-Speaker Audio

Meta presented Muse Voice Transcribe, the first real-time audio perception model developed by Meta Superintelligence Labs. [1] The artificial intelligence transcribes words as a person speaks and can identify who said what among more than 20 speakers. [1] It also detects when an intervention ends, integrating functions that typically require separate systems into a single tool. [1]

The model is trained on more than 70 languages and recognizes language changes within the same sentence. [1] For its initial launch, Meta validated 25 languages, including Chinese, French, Hindi, and Spanish. [1] The artificial intelligence also has the capacity to pause in order to continue the transcription later. [1]

By integrating transcription, speaker identification, and turn-detection into one tool, Meta targets the difficulty of processing audio with more than 20 distinct speakers. This system enters a market alongside Google's Gemini 3.5, which recently launched with the ability to transcribe over 85 languages, but Muse Voice Transcribe attempts to distinguish itself by processing mid-sentence language switches and managing high-volume multi-speaker environments in real time. What the sources don't address: How the model's accuracy and latency degrade when scaling up to the 30-voice limit shown in its demonstration.