Meta Releases 'Muse Voice Transcribe,' a Real-Time Speech Transcription Model

Meta has released a new AI model called 'Muse Voice Transcribe' that converts speech to text in real time. Developed by Meta Superintelligence Labs, the model integrates three functions—speech recognition, speaker identification, and end-of-utterance detection—into a single processing pipeline. The public API is priced at $0.18 per hour of processed audio.

The speech recognition field has recently seen growing demand for real-time processing in meeting systems and call centers, with multiple companies competing to offer their respective services. Speechmatics claims to identify up to 50 speakers by default and up to 100 with configuration changes, while Amazon Transcribe supports up to 30 speakers including streaming. While Muse Voice Transcribe's support for '20 or more' speakers is not the industry maximum, it positions itself at the upper tier of the market.

Meta highlights several key features of Muse Voice Transcribe: support for audio longer than one hour, processing of conversations with mixed languages (multilingual code-switching), a 'biasing' function that prioritizes recognition of specific languages or keywords, and speaker identification without post-processing. The model was trained on over 70 languages, with 25 languages receiving intensive validation at the time of initial release.

Technically, audio is divided into chunks of 80 milliseconds (12.5 times per second), with each chunk converted into a format called 'soft tokens.' The model continuously decides whether to wait for more audio or output text. Meta refers to this approach as 'adaptive latency'—when speech is unclear, the model delays its decision and, once sufficient context is available, rapidly finalizes the text. This behavior has been trained using reinforcement learning to maintain a balance between transcription accuracy and processing latency.

Speaker identification (diarization) is the technology that automatically determines 'who said what and when' from audio data. It is more challenging than basic transcription and has traditionally been performed as a separate post-processing step. By integrating speaker identification into the model itself, Muse Voice Transcribe completes the entire pipeline in real time, marking a significant departure from conventional architectures.

The importance of this feature becomes clear when viewed through the lens of the 'trustworthiness' of AI-generated records. For example, even if a meeting recording system correctly transcribes what was said, incorrectly attributing a statement to the wrong speaker can obscure accountability for decisions and agreements. Similar issues can arise in call center compliance management or AI agents operating in multi-party environments. From this perspective, embedding speaker identification at the core of real-time processing has substantial practical value.

Meta's pricing of $0.18 per hour appears competitive for enterprise speech AI services. For developers building systems that require continuous audio processing—such as meeting support tools, call analytics, and live assistants—the ability to handle transcription, speaker identification, and multilingual support through a single API at low cost could significantly influence adoption decisions. As competition in speech AI intensifies not only on features but also on price, it remains to be seen how Muse's arrival will affect pricing levels in this market segment.