Meta releases Muse Voice Transcribe, a real-time speech transcription model with built-in speaker diarization and sentence detection

Meta's Superintelligence Labs shipped Muse Voice Transcribe, a real-time speech-to-text model that also detects sentence boundaries and distinguishes 20+ speakers in one system. It processes audio in 80ms chunks, supports 70+ languages (25 tested in depth), handles code-switching, and is generally available now via Meta AI, Muse Code and the Meta Model API at $0.18 per hour ($3 per 1,000 audio minutes). Independent benchmarks from Artificial Analysis reportedly show 3.1% word error rate at 0.16s latency, ahead of ElevenLabs (3.6%) and AssemblyAI (4.0%).

This is a real, priced, generally-available product that developers and enterprises can procure today for live transcription, captioning, meeting tools or voice agents, so it's immediately actionable for anyone comparing streaming speech-to-text vendors. The combination of low latency, competitive accuracy and low cost pressures existing providers (OpenAI, ElevenLabs, AssemblyAI, Cartesia, Deepgram) on price and forces a re-evaluation of build-vs-buy decisions for speaker-attributed transcription. That said, the impact is bounded to a specific technical niche rather than a shift in AI capability overall.

The framing that this is "the foundation for AI assistants that never stop listening" and Zuckerberg's "personal superintelligence" narrative are vendor-supplied speculation layered onto what is otherwise a concrete, incremental product release. Streaming ASR with diarization is already a crowded, competitive commodity market with comparable recent launches from several rivals, so this is best read as a strong price and accuracy move, not a breakthrough. The extraction also lacks primary source links, so the benchmark figures should be verified against Artificial Analysis directly rather than taken on trust.

1. 01Independent, reproducible benchmark comparisons (not just Artificial Analysis) confirming the claimed 3.1% WER and 0.16s latency across diverse audio conditions, not just the tested 25 languages
2. 02Developer adoption signals: usage volume, published case studies, or third-party integrations citing Muse Voice Transcribe within 3-6 months of launch
3. 03Competitor pricing or capability responses from OpenAI, ElevenLabs, AssemblyAI, Deepgram and Cartesia in the following quarter, indicating whether this actually shifted the competitive floor