Meta Releases MuseVoiceTranscribe: Real-time Audio Perception Model for 20+ Speakers
智东西 · wechat · 2026-09-02
Meta has released MuseVoiceTranscribe, its first real-time audio perception model, aiming to upgrade speech transcription into a real-time sensing layer for AI systems. Key capabilities include:
- Multilingual & Long Audio: Supports 70+ languages (25 verified), natively handles audio over 1 hour, and enables seamless code-switching (e.g., Chinese-English mixing).
- Multi-speaker Diarization: Natively supports diarization for 20+ speakers without post-processing, accurately identifying speaker turns.
- Streaming & Adaptive Latency: Uses an autoregressive multimodal architecture processing audio in 80ms chunks. It leverages Reinforcement Learning (RL) for "adaptive latency," allowing the model to dynamically decide listening duration per word to balance speed and accuracy optimally.
- Voice Activity Detection: Introduces special tokens to mark speech onset and endpoints.
The model is integrated into Meta AI, MuseCode, and MetaModelAPI, priced at approx. $3 per 1,000 minutes.
More from Multimodal
- H3 360° Orbit Camera LoRA for MiniMax Open-Sourced, Generates Smooth 5.2s Shots — merica420_69 · 2026-10-18
- 'Tesla in Pakistan' AI video made with Seedance 2.5 goes viral — SimplyAnnisa · 2026-10-18
- Borges's nine stories turned into walkable 3D worlds with AI model Fable 5.1 — emollick · 2026-10-18
- Midjourney cozy couple photo: full prompt shared for killing the AI plastic look — tisch_eins · 2026-10-18
- Higgsfield playbook: scene templates, paid creator network, $20-to-enterprise funnel — sujingshen · 2026-10-18
- Why did Anima flop? Illustrious still dominates the anime model scene — Front-Side-6346 · 2026-10-18