Meta has launched Muse Voice Transcribe, its first real-time speech recognition model. Developed by Meta Superintelligence Labs, the system converts speech to text, separates speakers, and determines whether the conversation has ended. The model can process recordings longer than one hour and calls with more than 20 speakers without additional processing. Trained in more than 70 languages, 25 of them underwent extensive validation before the first version; the model can recognize language switching and proper names based on context. The “adaptive latency” system, which processes conversations in 80-millisecond segments, balances speed and accuracy. By distinguishing the beginning and end of a conversation, it determines response timing in voice applications. Meta announced that the model ranked first in the Artificial Analysis rankings as of September 1. The technology has been added to Meta AI desktop dictation and Muse Code. The API costs 3 dollars per thousand minutes, or approximately 18 cents per hour. Its integration into WhatsApp, Instagram, and Messenger is unknown.
Why it matters
The practical impact of the development lies not only in enabling voice applications to transcribe speech, but also in allowing them to process who is speaking and when the interaction has ended. This is particularly relevant to developers working with long recordings, multi-participant conversations, and conversations involving multiple languages; its ability to distinguish language switches and identify proper names from context could also provide additional value in these use cases. Real-time processing and adaptive latency allow applications to determine when to generate responses based on the model’s conversational flow. While the disclosure of the API price enables those wishing to use the technology to calculate costs in advance, whether the model will be added to WhatsApp, Instagram, and Messenger remains an open question in terms of its impact on the user experience.
Background
Meta is not a new name in the FikirPilot archive: over the past 90 days, we have published 4 articles mentioning the name, the latest dated August 30, 2026.