Meta's new real-time speech-to-text model tops OpenAI and Google on accuracy benchmarks

Meta's Superintelligence Labs unveiled Muse Voice Transcribe this week, a specialized real-time speech recognition system that demonstrates strong performance across multiple benchmarks compared to rivals like OpenAI, Google, and ElevenLabs.

Performance and capabilities#

The model achieves a 3.1% word error rate on English transcription tasks, according to Artificial Analysis's AA-WER Streaming benchmark—outpacing Cartesia Ink-2 (3.4%), ElevenLabs' Scribe v2 Realtime (3.6%), OpenAI's GPT Live Transcribe (3.9%), and Google's Gemini 3.5 Transcribe Live (4%).

Beyond accuracy, Muse Voice Transcribe handles several practical demands: it can distinguish among 20+ speakers, supports over 70 languages (with 25 extensively validated), and processes multilingual conversations where speakers switch languages mid-stream. It also manages extended sessions lasting more than an hour. In speaker diarization—identifying who is speaking—the model posts a 17.5% error rate on standard benchmarks, still leading the pack despite remaining room for improvement.

How it works#

Muse Voice Transcribe uses an autoregressive multimodal architecture derived from Meta's Muse Spark family. The system ingests audio as 80-millisecond chunks, which it compresses into individual tokens. At each chunk, the model decides whether to output a text token or insert a placeholder for the next audio segment.

This design gives the model control over latency. Simpler words receive transcription nearly immediately, while challenging passages gain additional context before commitment—a balance Meta calls "adaptive delay." This tradeoff is learned through reinforcement learning by combining word error rate rewards and delay rewards multiplicatively, rather than simply adding them.

The same token-switching approach handles speaker detection.

Availability and economics#

The model is accessible through Meta Model API, Meta AI for Mac, and Muse Code, priced at $3.00 per 1,000 audio-minutes ($0.18 per hour). Unlike Meta's open-weight Glimmer models, Muse Voice Transcribe remains proprietary.

Market context#

Real-time transcription has become intensely competitive in recent months, with OpenAI, Google, xAI, and Alibaba launching streaming variants alongside established players. While Meta's benchmark lead is notable, maintaining that advantage in a rapidly evolving space will require continued investment—something the company is well-positioned for given its integration needs across glasses and desktop products.