Is Meta's Muse Voice Transcribe Really SOTA?

Verified

Meta Superintelligence Labs launched Muse Voice Transcribe on September 1, 2026, with CEO Mark Zuckerberg calling it 'SOTA in streaming speech-to-text.' Independent benchmark firm Artificial Analysis confirms the model ranks #1 on its AA-WER Streaming test at 3.1% word error rate, ahead of Cartesia Ink-2 (3.4%), ElevenLabs Scribe v2 (3.6%), OpenAI's GPT Live Transcribe (3.9%), and Google's Gemini 3.5 Transcribe Live (4%). The lead is real but narrow, and the test measures English only.

The claim

"Meta's Muse Voice Transcribe model is state-of-the-art (SOTA) in streaming speech-to-text, beating models from OpenAI, Google, Cartesia, and ElevenLabs." — Meta AI Research, 2026-09-01

Supporting evidence

Independent benchmark tracker Artificial Analysis confirms Muse Voice Transcribe takes the #1 spot for final-transcript accuracy on its AA-WER Streaming test, at 3.1% word error rate — ahead of every other tracked model. Artificial Analysis (X/Twitter)
Meta's own research announcement details the model's streaming architecture, training data, and launch benchmarks. Meta AI Research
A fact-checked breakdown corroborates the 3.1% WER figure against each named competitor's score, confirming Meta's benchmark lead is real. SQ Magazine

Contradicting evidence

The AA-WER Streaming test that backs the SOTA claim covers English speech only; Meta's 70-plus-language training claim falls outside what the leaderboard actually measures. SQ Magazine
Meta's lead over the next-closest model (Cartesia Ink-2) is only 0.3 percentage points, and Meta's speaker-diarization accuracy trails at a 17.5% error rate — well short of solved, despite Meta also leading that metric. SQ Magazine

Frequently asked questions

What benchmark supports Meta's 'SOTA' claim?

Artificial Analysis's AA-WER Streaming test, which scores word error rate for real-time transcription. Muse Voice Transcribe recorded 3.1% WER, the lowest of any model the firm tracks.

Does the SOTA claim cover all 70-plus languages Meta trained on?

No. The benchmark backing the claim is English-only. Meta says it trained on more than 70 languages but had extensively validated only 25 of them at launch.

Is Muse Voice Transcribe also state-of-the-art at identifying individual speakers?

Not by the same margin. Meta leads public speaker-diarization benchmarks, but at a 17.5% error rate — a level industry coverage describes as still far from solved across the field.

Reviewed by the AI Ledger editorial team · Last updated 2026-09-03. Verdicts follow our published methodology; spot an error? tell us and we'll re-review it.