Meta has unveiled Muse Voice Transcribe, an innovative real-time audio perception model that integrates multilingual, streaming transcription into Meta AI for Mac and Muse Code. This technology is also accessible to developers via the Meta Model API, offering a robust solution for diverse transcription needs.
The Muse Voice Transcribe system, developed by Meta Superintelligence Labs, incorporates automatic speech recognition with speaker diarization and intelligent endpoint detection. This means it can transcribe spoken words as they occur, differentiate between more than 20 distinct voices in a recording, and accurately identify when a speaker has concluded their utterance, all without the necessity of post-processing. The model's training spans over 70 languages, with 25 fully validated at launch, and it supports audio durations exceeding an hour. A notable feature is its native code-switching capability, allowing seamless transitions between languages or within sentences, while options for language, keyword, and context biasing further refine its recognition accuracy.
Instead of adhering to a fixed speed-accuracy balance, Muse Voice Transcribe employs an "adaptive delay" mechanism, which dynamically adjusts the listening duration before committing a word to text. This allows for swift processing of straightforward speech and more extensive audio context utilization for complex words. Currently, the model holds the top position on the Artificial Analysis streaming speech-to-text leaderboard. Available through the Meta Model API at a rate of $3 per 1,000 audio-minutes, it already powers dictation functionalities in Meta AI for Mac and Muse Code, enabling Mac users to dictate into any application by simply holding the Fn key.
This advancement represents a significant stride in speech-to-text technology, offering enhanced precision and efficiency in real-time transcription across multiple languages and contexts. It underscores the potential of AI to seamlessly integrate into daily workflows, fostering greater productivity and accessibility for users and developers alike.
