Muse Voice Transcribe Benchmarks

Meta reported that Muse Voice Transcribe ranked first on Artificial Analysis streaming speech-to-text and on public diarization benchmarks. The announcement explicitly dates model inclusion and rankings to September 1, 2026.

• Treat the rank as a launch-day snapshot.
• Link the current leaderboard next to any numerical comparison.
• Keep streaming and non-streaming results in separate tables.

The 3.1% word error rate figure should be attributed to Artificial Analysis' September 1, 2026 streaming speech-to-text snapshot, where Muse Voice Transcribe was reported at roughly 0.16 seconds after end of speech. It is third-party leaderboard evidence, not a permanent official accuracy guarantee for every language, microphone, accent, or domain.

• Call it an Artificial Analysis snapshot, not an evergreen Meta benchmark.
• Keep the date, streaming mode, and end-of-speech latency next to the number.
• Retest on your own audio before using it for procurement or product claims.

Streaming transcription is a two-axis problem: a system can wait longer to improve its final words, or respond sooner with less context. Meta's adaptive-delay training is designed to move along that trade-off word by word.

Word error rate measures transcription, while diarization error rate measures who spoke when. A trustworthy comparison identifies the datasets, speaker conditions, streaming mode, and whether post-processing was allowed.

We do not present fixed 3.1% WER or latency as official facts without the Artificial Analysis date and test context. The $0.18/hour API price belongs to Meta Model API pricing materials and should not be mixed with this site's browser demo.