Meta Superintelligence Labs Launches Muse Voice Transcribe, Claiming Streaming ASR SOTA

On September 2, Meta Superintelligence Labs released Muse Voice Transcribe, its first real-time audio perception model, claiming SOTA on streaming speech-to-text. Per Artificial Analysis, it ranks first in final transcription accuracy on the AA-WER Streaming benchmark with 3.1% WER at very low latency. Positioned as Meta's first step toward making transcription a real-time perception layer for AI systems, it was announced by finkd and widely amplified by independent observers such as @rohanpaulai and @bowenc0221, with Alexandr Wang also promoting it (attributing the launch to Scale AI). It is available via the Meta model API, in the Meta desktop app (e.g., macOS Meta AI) and Muse Code, at roughly $3 per 1,000 minutes of audio per @SumitGup.

Confirmed

• Core capabilities confirmed by Meta's official @AIatMeta account: low-latency streaming ASR, 20+ speaker diarization and endpointing, natively integrated in a single model
• Trained on 70+ languages with 25 verified at launch, supporting seamless intra-sentence code-switching
• Handles audio over one hour long; supports language/keyword/context biasing for accuracy
• @AIatMeta details a Pareto-optimal speed-accuracy tradeoff via adaptive latency (fast output on easy words, longer waits on hard ones), trained with reinforcement learning combining WER and latency rewards
• Per @testingcatalog, the model is live on the Meta model API with a zero data retention tier
• Per @SumitGup, cost is about $3 per 1,000 minutes of audio

Unconfirmed

• Attribution of the launch: Alexandr Wang and several reposters credit Scale AI, while Meta's official @AIatMeta and outlets like Zhidxing attribute it to Meta Superintelligence Labs; the discrepancy, likely tied to branding after Scale AI's absorption into Meta, is not resolved in the posts

Why it matters

• Natively integrating streaming ASR, diarization and endpointing in one model eliminates traditional pipeline assembly, with direct applications in real-time translation, meeting notes and voice assistants
• 20+ speaker diarization and intra-sentence code-switching target long-standing weaknesses in real-world multilingual, multi-speaker settings
• The adaptive-latency RL training recipe offers a new approach to the speed-accuracy tradeoff in streaming ASR
• Low pricing ($3 per 1,000 minutes) plus a zero data retention tier lower cost and privacy-compliance barriers, signaling Meta's push into commercial API services

2026-09-02 ~ 2026-09-02 · 18 related posts

• Meta releases Muse Voice: SOTA real-time speech-to-text model for 70+ languages — bowenc0221 · 2026-09-02
• Meta's Muse Voice Transcribe hits API with zero-data-retention tier — testingcatalog · 2026-09-02
• [source] Meta's Muse Voice Transcribe Balances Speed and Accuracy with Adaptive Delay — AIatMeta · 2026-09-02
• [source] Meta Releases Muse Voice Transcribe: Real-Time ASR with 20+ Speaker Diarization — AIatMeta · 2026-09-02
• Meta Releases Muse Voice Transcribe: Real-Time Streaming Speech Model with Diarization — bowenc0221 · 2026-09-02
• MSL launches Muse Voice Transcribe, a streaming audio model with real-time ASR and diarization — bowenc0221 · 2026-09-02
• Scale AI Releases Muse Voice: SOTA Streaming Speech-to-Text Model — alexandr_wang · 2026-09-02
• [source] Meta releases Muse Voice Transcribe with top streaming accuracy — ArtificialAnlys · 2026-09-02
• MSL Releases Muse Voice: SOTA Real-Time Transcription Model — bowenc0221 · 2026-09-02
• Meta releases Muse Voice Transcribe: Real-time ASR with 20+ speaker diarization — shuchaobi · 2026-09-02
• Meta releases real-time audio model at $3 per 1000 minutes — SumitGup · 2026-09-02
• Meta Launches Muse Voice Transcribe: Streaming STT With 3.1% Word Error Rate — rohanpaul_ai · 2026-09-02

6 near-duplicate retellings: bowenc0221 · rohanpaul_ai · 智东西 · aigclink · jffwng · alexandr_wang