Meta Superintelligence Labs on Tuesday launched Muse Voice Transcribe, a real-time speech recognition model that leads the Artificial Analysis AA-WER Streaming speech-to-text accuracy benchmark with a 3.1% word error rate in English1.
The model outperforms Cartesia Ink-2 (3.4%), ElevenLabs Scribe v2 Real-time (3.6%), GPT Live Transcribe (3.9%), and Gemini 3.5 Transcribe Live (4%) on the same benchmark. Meta's lab describes Muse Voice Transcribe as its first real-time audio perception model2.
Muse Voice Transcribe can distinguish among more than 20 speakers and has been trained on over 70 languages, with 25 of them extensively verified. The model supports multilingual speakers switching languages mid-conversation and can handle sessions longer than an hour.
The model is available via the Meta Model API, Meta AI for Mac, and Muse Code, priced at $3.00 per 1,000 audio minutes ($0.18 per hour). Meta will not release the open weights, a Meta spokesperson told The New Stack. The closed-weight decision contrasts with Meta's approach to its Muse Glimmer models.
Architecturally, Muse Voice Transcribe is an autoregressive multimodal model from the Muse Spark family. Audio arrives in 80-millisecond chunks at 12.5 per second, with each chunk compressed into a single soft token. At every chunk, the model either emits a text token or a special "next audio" placeholder; when audio stops, an empty audio token signals the model to flush remaining text.
Meta calls the latency mechanism "adaptive delay": the model controls how much audio it consumes before committing to a word, giving difficult words more context while transcribing easy words almost immediately. The tradeoff between word error rate and delay is learned during a reinforcement learning phase, where the two reward signals are multiplied rather than added. A similar mechanism handles speaker detection.
Speaker diarization remains a weak point across the field. Muse Voice Transcribe carries a 17.5% error rate across several standard benchmarks for recognizing distinct speakers.
The AA-WER benchmark that Muse Voice Transcribe leads applies only to English speech. Meta recently shipped Muse Spark, another speech-to-text capable model, though it does not specialize in real-time transcription. OpenAI, Google, xAI, and Alibaba have all shipped streaming models within weeks of each other this summer, according to The New Stack.
ANALYSIS The closed-weight release marks a departure from Meta's usual open-model strategy, particularly given the explicit contrast with Muse Glimmer. Pricing at $0.18 per hour positions the model at a concrete cost point for developers building always-on transcription into hardware products such as Meta's smart glasses and its Mac app, both of which depend on real-time speech processing.