Scribe v3 hits 74x real-time transcription: most accurate ASR but far slower than rivals

ArtificialAnlys · x · 2026-11-20

Artificial Analysis updated its speech-to-text leaderboard: Scribe v3 transcribes at 74x real time (up from Scribe v2's 58x) and is the most accurate model tested, but slower than MAI-Transcribe-2 (334x) and Zoom Scribe V2.1 Pro (188x), and faster only than Fun-Realtime-ASR-preview (21x). The AA-WER v2.2 index weighs 8 hours of audio across AA-AgentTalk (50%), VoxPopuli (25%), and Earnings22 (25%), covering 90 models with diverse accents and challenging acoustics.

Related event: ElevenLabs launches Scribe v3: WER drops to 1.95%, second only to Alibaba's Fun-ASR(5 posts)→

Original post →

More from Models

Models channel →