Meta's Superintelligence Labs has released Muse Voice Transcribe, a model designed to turn spoken words into text in real time. The system processes audio in 80-millisecond chunks, a speed that allows it to keep up with natural conversation. It can distinguish between different speakers and recognise when a sentence ends. Artificial Analysis has evaluated the model and claims it offers the most accurate streaming transcription at the lowest price in the current market. Meta sees this technology as a foundational component for personal AI agents that operate continuously through devices such as smart glasses.

The focus on low latency and speaker separation suggests a specific use case for always-on assistants. By detecting sentence boundaries, the model can trigger actions or responses without waiting for a pause in speech. This capability could enable more fluid interactions where the AI listens and acts in the moment, rather than waiting for a full query. The Decoder notes that Meta intends to integrate this model into hardware that records conversations, positioning the technology as a key enabler for future personal AI.