Meta's Superintelligence Labs on Tuesday released Muse Voice Transcribe, a real-time speech recognition system that achieves a 3.1% word error rate on English benchmarks, outperforming competitors including OpenAI's GPT Live Transcribe and Google's Gemini 3.5 Transcribe Live. The company describes the system as its first dedicated real-time audio perception model, built to handle multilingual conversations, distinguish between more than 20 distinct speakers, and process discussions lasting over an hour. Unlike Meta's recent Muse Glimmer models, this one won't be released with open weights, according to a company spokesperson.

On Artificial Analysis's AA-WER Streaming speech-to-text accuracy benchmark, Muse Voice Transcribe posted a 3.1% error rate, ahead of Cartesia Ink-2 at 3.4%, ElevenLabs' Scribe v2 Real-time at 3.6%, GPT Live Transcribe at 3.9%, and Gemini 3.5 Transcribe Live at 4%. The model has been trained on more than 70 languages, with 25 of those languages extensively verified, and it can handle cases where speakers switch languages mid-conversation. For speaker recognition in real-time scenarios, the system records a 17.5% error rate across multiple standard benchmarks, leading the field even though all models still struggle considerably in this area. The service is now available through the Meta Model API at $3.00 per 1,000 audio minutes, which works out to $0.18 per hour, and it's also integrated into Meta AI for Mac and Muse Code.

Meta describes the model as an autoregressive multimodal system from the Muse Spark family, and the company explains that audio arrives in 80-millisecond chunks, compressed into single soft tokens at a rate of 12.5 per second. At each chunk, the model makes a decision: it either outputs a text token or a special placeholder that gets replaced with the next audio chunk. When audio stops, an empty audio token tells the model no more input is coming, prompting it to output any remaining text. The model controls how much audio it consumes before committing to a word, creating what Meta calls "adaptive delay"—difficult words receive more context while simple words get transcribed almost instantly. This tradeoff is learned during reinforcement learning, where word error rate and delay rewards are multiplied rather than added.

Meta notes that real-time transcription has become one of the most competitive segments of the AI market this summer, with OpenAI, Google, xAI, and Alibaba all shipping streaming models within weeks of each other, joining existing specialist providers. The report acknowledges that a 0.3-point benchmark advantage won't last long in such a crowded field. But Meta points to a built-in motivation to keep advancing: every product the company prioritizes, from its smart glasses to the Mac application, depends on this technology working as effectively as possible. For companies betting on voice interfaces becoming standard across consumer hardware, the difference between acceptable and exceptional transcription will increasingly determine which platforms users trust with ambient listening.