ASR systems reaching human-level accuracy in legal transcription
#1OpenAI's Whisper Large V3 Turbo, Google's Universal Speech Model, and Deepgram Nova-2 have achieved word error rates below 5% on clean English audio benchmarks. Legal-domain fine-tuned models from Verbit and Rev are further closing the gap, with some claiming 97%+ accuracy on deposition audio. The remaining accuracy gap narrows with each model generation, typically improving 1-2% annually.