Skip to content
NewsModel release

Meta Superintelligence Labs Ships Muse Voice Transcribe, Combining Real-Time ASR, Diarization, and Endpointing in One Model

· by Pondero Newsdesk

The short version

Meta released Muse Voice Transcribe on September 1, 2026, the first real-time audio perception model to handle streaming transcription, speaker identification for 20-plus voices, and sentence detection in a single end-to-end architecture at $3.00 per 1,000 audio minutes.

Meta Superintelligence Labs Ships Muse Voice Transcribe, Combining Real-Time ASR, Diarization, and Endpointing in One Model

Before Muse Voice Transcribe, building a production voice pipeline meant stitching together three systems: one for transcription, one for speaker identification, and one for sentence boundary detection. Meta collapsed that into a single model on September 1, 2026.

What

Meta Superintelligence Labs released Muse Voice Transcribe as the first real-time audio perception model in its Muse family, per Meta's research blog. The model processes audio in 80-millisecond chunks at 12.5 Hz, performing streaming ASR, speaker diarization for up to 20 simultaneous voices, and endpointing in one end-to-end architecture. No separate post-processing step is required after transcription.

The underlying design is autoregressive: each audio chunk becomes a soft token, and the model decides at each step whether to keep listening or emit text. A reinforcement-learning-trained "adaptive delay" policy adjusts per-word how long the model waits before committing, trading latency for accuracy dynamically. The result, per MarkTechPost's technical breakdown, is a 3.1% streaming word error rate, ranking first on Artificial Analysis's streaming speech-to-text benchmark as of September 1, 2026. Its diarization error rate across standard benchmarks sits at 17.5%, compared to competitors ranging from 21.1% to 28.6%.

The model handles recordings exceeding one hour, supports up to 20-plus concurrent speakers, and switches between languages natively within a single sentence. It was trained on more than 70 languages, with 25 extensively validated at launch per MarkTechPost.

Pricing is $3.00 per 1,000 audio minutes ($0.18 per hour). No weights were released. The model ships as muse-voice-transcribe-1.0 on the Meta Model API and also powers voice dictation inside Meta AI for Mac and Muse Code, Meta's AI coding environment.

Why it matters

The consolidation from three pipeline components to one changes the operational math for teams running real-time transcription at any scale. Three separate models meant three latency budgets, three failure modes, and coordination at every hand-off between systems. A single end-to-end model with native diarization simplifies that architecture considerably, and the 80ms chunk size keeps the interactive feel for voice dictation use cases.

At $0.18 per audio hour, the pricing is competitive with ElevenLabs' Scribe tier and accessible for call-center analytics, meeting transcription, and live captioning. The benchmark numbers are vendor-reported via Artificial Analysis, so independent confirmation across diverse accents and noise conditions still matters before committing. But a 3.1% WER at 80ms chunk latency, if it holds in production, would represent a genuine accuracy-latency improvement over the current Pareto frontier among hosted streaming transcription APIs.

The absence of open weights is a hard constraint for any team with data-residency or on-premise requirements. Every inference routes through Meta's API.

What to watch next

Two signals are worth tracking. First, whether independent benchmarking confirms the Artificial Analysis numbers across accent and noise profiles beyond the demo scenarios Meta showed. Second, whether platform partnerships follow: Meta's framing positions this model as foundational for personal AI assistants and AR glasses. If Apple or Microsoft embed Muse Voice Transcribe into their productivity stacks, pricing and API terms could shift from the current developer-first model.

Sources