AudioBench - People's Speech WER: leaderboard
AudioBench People's Speech word error rate slice measuring automatic speech-recognition transcription quality, where lower scores indicate fewer errors.
Metric: WER (%) (lower is better). Source: github.com. 9 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | cascade_whisper_large_v3_llama_3_8b_instruct | 14.54 | |
| 2 | whisper_large_v3 | 14.6 | |
| 3 | cascade_whisper_large_v2_gemma2_9b_cpt_sea_lionv3_instruct | 20.14 | |
| 4 | MERaLiON-AudioLLM-Whisper-SEA-LION | 21.05 | |
| 5 | Qwen2-Audio-7B-Instruct | 21.65 | #880 |
| 6 | SALMONN_7B | 23.7 | |
| 7 | Qwen-Audio-Chat | 31.42 | #1244 |
| 8 | seallms_audio_7b | 36.98 | |
| 9 | WavLLM_fairseq | 37.92 |
Interactive version: theaggregate.ai/benchmark?slug=audiobench-people-s-speech-wer · How It Works · Data refreshed daily, snapshot 2026-10-20.