ElevenLabs and Google Lead Artificial Analysis' Updated Speech-to-Text Benchmark

2026-03-02

Artificial Analysis has officially unveiled version 2.0 of its AA-WER speech-to-text benchmark. Topping the charts is ElevenLabs' Scribe v2, which achieved an impressive word error rate (WER) of just 2.3%. Close behind are Google's Gemini 3 Pro at 2.9% and Mistral's Voxtral Small at 3.0%. Google's Gemini 3 Flash (3.1%) and the previous generation ElevenLabs Scribe v1 (3.2%) follow closely. Notably, Google did not undergo specific training for transcription tasks; instead, these robust results stem from Gemini's inherent general multimodal capabilities. OpenAI's popular open-source model, Whisper Large v3, sits in the middle tier with 4.2%, while Alibaba's Qwen3 ASR Flash (5.9%), Amazon's Nova 2 Omni (6.0%), and Rev AI (6.1%) round out the lower end of the rankings.

These performance trends remain consistent in the separate AA-AgentTalk evaluation designed for voice assistants. ElevenLabs' Scribe v2 (1.6%) and Google's Gemini 3 Pro (1.7%) maintain a significant lead, with AssemblyAI's Universal-3 Pro securing third place at 2.3%.

RECENT AI NEWS

RECENT AI TOOLS