Meta launches real-time audio perception AI model for multi-speaker transcription

Meta has introduced Muse Voice Transcribe, its first real-time audio perception model, capable of transcribing speech from over 20 speakers simultaneously. The model integrates streaming automatic speech recognition, speaker separation, and endpoint detection, allowing for immediate text output as users speak. Developers can access the model via the Meta Model API, with pricing at $3 per 1,000 minutes of audio, equivalent to approximately 1.2 yuan per hour. The model supports over 70 languages, with 25 verified at launch, and can handle audio longer than an hour. It also allows for natural language switching within or between sentences. Meta claims the model ranks first in the Artificial Analysis streaming speech-to-text benchmark as of September 1, 2026. The model employs an adaptive delay mechanism to balance real-time performance and accuracy, dynamically adjusting the amount of audio processed before outputting results. For clearer speech, the system provides faster transcriptions, while more complex or unclear speech is processed with additional context. This approach aims to enhance both speed and precision. The model’s ability to separate speakers and detect speech endpoints makes it suitable for real-time transcription, meeting, and call captioning applications.