Meta has launched Muse Voice Transcribe. The model covers more than 70 languages worldwide. Native support includes five major Indian languages: Hindi, Tamil, Telugu, Kannada, and Malayalam. Meta Superintelligence Labs built it as the company’s first real-time audio perception model. One system handles live transcription, speaker separation, and endpointing. It also follows code-switching and crowded conversations.
The tool writes speech as people talk. It does not wait for the full recording to finish. It labels who is speaking. It marks when a turn starts and stops. Those tasks run on the incoming stream. No separate cleanup step is required after the fact.
Meta says the model can tell apart more than 20 speakers in one file. It can work on audio longer than an hour. Long meetings therefore do not need to be chopped into short clips first. Audio arrives in 80-millisecond chunks. The model decides when it has heard enough to lock each word. Harder words wait longer. Simpler words commit sooner. Reinforcement learning guides those delays. The aim is fewer errors without extra lag.
Training covered more than 70 languages. Twenty-five of them received extra checks for the first release. Indian language support is part of that set. Speakers can switch languages mid-conversation. They can even switch inside a sentence. Users do not have to tap a new language setting each time. Biasing tools use language, keywords, and wider context. Those cues help the model catch names and domain terms from the surrounding talk.
Developers can call the model through the Meta Model API. Meta AI for Mac and Muse Code already use it for dictation. API pricing is $3, or about Rs 300, per 1,000 audio minutes. That works out to about $0.18, or roughly Rs 17, per hour.
In short, Meta is selling live speech-to-text that tries to follow mixed-language rooms and many speakers at once. Indian language coverage is a headline feature for this market. Accuracy on noisy calls and heavy code-switching will decide how useful the API becomes.
