Meta launches Muse Voice Transcribe, a real-time audio model billed by the minute

Meta has launched its first real-time audio model, Muse Voice Transcribe. Developers can use it through the Meta Model API, priced at $3 for every 1,000 minutes of audio. The model combines streaming speech recognition, speaker separation, and endpoint detection, so it can output text as you speak without waiting for the full recording to finish.

Muse Voice Transcribe can separate different speakers in recordings with more than 20 people, and it automatically detects when someone stops speaking to set boundaries. It was trained on over 70 languages, with 25 verified at launch. It also handles audio longer than an hour and allows natural language switching within or between sentences.

To balance speed and accuracy, Meta added an adaptive delay system that decides how much extra audio to wait for each word before giving the result. Meta says the model ranks first on the Artificial Analysis streaming speech-to-text chart as of September 1, 2026.