Meta has introduced Muse Voice Transcribe for real-time audio processing. The model converts speech into text during conversations without waiting for a recording to finish, while also identifying different speakers and detecting when they start and stop speaking.
Muse Voice Transcribe was developed by Meta Superintelligence Labs. The model has been trained on more than 70 languages and supports Hindi, Tamil, Telugu, Kannada and Malayalam among the Indian languages.
Before the initial release, Meta tested the model on 25 languages. It can also handle language switching during conversations, including situations where the language changes within the same sentence, without requiring users to manually change language settings.
The system supports Language, Keyword and Context Biasing to help identify words based on the context of a conversation.
According to Meta, Muse Voice Transcribe can distinguish between the voices of more than 20 people in a single recording. It can also process audio longer than one hour, reducing the need to divide long conversations into shorter clips.
The model processes audio in small segments of about 80 milliseconds. It then determines whether enough information is available to accurately transcribe a word. Words that are easier to identify can be processed quickly, while the model waits for additional audio when a word is unclear.
Meta said Reinforcement Learning is used in this process to help the model determine the appropriate timing for transcription during speech and reduce errors.
Meta has made Muse Voice Transcribe available through the Meta Model API. Developers can use it to add speech transcription features to their applications for real-time audio use cases.
The model is also being used for dictation in Meta AI for Mac and Muse Code.
The API is priced at $3 (around ₹300) per 1,000 audio minutes. At this rate, one hour of audio costs approximately $0.18 (around ₹17).
