← Back to feed

Cohere Transcribe Hits 1.46% WER Clean, 2.5x Worse Than Claimed on Real Audio

AI · Studies · 2026-10-19 00:48

Benchmarks show Cohere's new transcription model achieves 13.65% word error rate on real audio, far exceeding its claimed 5.42%, performing well only on clean speech data.

Source: inferencebench.io

Auto-generated summary for reference only. RSS