Meta releases Muse Voice Transcribe and Muse Spark 1.3
Both models are available through the Meta Model API and Muse Code, offering streaming speech transcription and more efficient agentic task execution.
- Muse Voice Transcribe provides real-time speech-to-text with diarization for over 20 speakers, multilingual code-switching, and context biasing.
- Muse Voice Transcribe processes audio in 80ms chunks, using reinforcement learning-trained adaptive delay to balance speed and accuracy per word.
- Muse Spark 1.3 consumes approximately 25% fewer tokens and 20% fewer tool calls than Muse Spark 1.2 on internal benchmarks.
- Muse Spark 1.3 handles longer-horizon tasks in a single thread, asks clarifying questions, and seeks confirmation before consequential actions.
- A max reasoning option for Muse Spark 1.3 will roll out once safety testing is complete.
Developers building on Meta's developer tools gain access to real-time audio transcription alongside more token-efficient agentic coding workflows.
