On September 1, 2026, Meta announced Muse Voice Transcribe, the first real-time audio perception model developed by its Superintelligence Labs. The model delivers streaming automatic speech recognition, speaker diarization for 20-plus speakers, and endpoint detection in a single autoregressive architecture.
Muse Voice Transcribe processes audio in 80-millisecond chunks at 12.5 Hz, converting each segment into a soft token. The model decides at each step whether to continue listening or emit text, using an adaptive delay mechanism trained with reinforcement learning to balance word error rate against latency. Meta reports a final-transcription streaming word error rate of 3.1 percent on the Artificial Analysis benchmark, placing it ahead of seven competing systems that range from 3.4 to 4.0 percent. On diarization, the model achieves an average error rate of 17.5 percent across AMI-IHM, AMI-SDM, and VoxConverse datasets, compared with 21.1 to 28.6 percent for five other systems.
What's new
- Streaming ASR, diarization, and endpointing in one model — no separate batch pass required.
- 25-plus languages with seamless code-switching inside a single sentence.
- Context and keyword biasing for proper nouns, contacts, and domain terms without fine-tuning.
- Adaptive delay that dynamically adjusts per word based on difficulty.
- Speaker attribution for 20-plus concurrent speakers with turn-aware token tags.
- Pricing at $0.18 per hour ($3.00 per 1,000 minutes) via the Meta Model API.
The model is already powering dictation in the Meta desktop app and voice input in Muse Code, Meta's AI-assisted development environment. Developers can access it through the Meta Model API using an OpenAI SDK-compatible client.
Why it matters
Real-time streaming transcription with integrated diarization and endpointing has been a persistent gap for voice interfaces that aim to handle natural, multi-speaker conversations. Most production systems still stitch together separate ASR, diarization, and voice-activity-detection models, which adds latency and compounds errors. By unifying these tasks in a single autoregressive model that emits special tokens for speaker turns and speech boundaries, Meta reduces the engineering overhead for applications such as meeting transcription, live captioning, and on-device dictation. The 20-plus speaker capacity also addresses a practical ceiling in current diarization benchmarks, which typically top out at fewer concurrent speakers.
Our take
Meta is positioning Muse Voice Transcribe as the sensory layer for "personal superintelligence" on future AI glasses — ears that work in messy, multilingual, overlapping conversation. The technical achievement is real: a single model that leads public streaming benchmarks while exposing a simple token-level API. The open question is whether the 17.5 percent diarization error rate on academic datasets translates to reliable speaker attribution in noisy, reverberant real-world settings, and whether the $0.18-per-hour pricing holds at scale for always-on wearable use cases.
