Meta releases Muse Voice Transcribe for real-time multilingual speech
Meta has released Muse Voice Transcribe, a real-time audio model developed by its Superintelligence Labs. It transcribes speech, identifies up to 20 speakers and marks sentence boundaries within one system. Incoming audio is processed in 80-millisecond segments, while the model decides how long to listen before producing each word. Meta says reinforcement learning taught it to give easy words shorter delays and difficult ones more processing time.
The model supports more than 70 languages, including conversations that switch languages mid-sentence. It can use language, keyword and context hints to improve recognition of names and other proper nouns. Meta says speaker attribution and sentence detection work during recordings longer than an hour without post-processing.
In Artificial Analysis testing conducted on September 1, 2026, Muse Voice Transcribe recorded a 3.1 percent English word error rate and produced the final transcript 0.16 seconds after speech ended. Meta offers it in Meta AI and Muse Code, and through the Meta Model API. Its price is $0.18 per hour, or $3 per 1,000 audio minutes, below the listed prices of competing services. Meta has not disclosed the model’s parameter count, training-data volume or audio sources, and is not releasing its weights.
Written by Sakhanda Wire AI from reporting by The Decoder. Facts, quotes and full context are in the original article below.