Fact · Transcription
5.81percent word error rate
In a 2025 challenge on non-native English, a fine-tuned Whisper system scored 5.81 percent word error rate when it labeled hesitations, against 6.47 percent when it removed them.
Checked against its source on October 17, 2026
In the source's own words
“Our challenge system achieved 6.47% WER (Pure) and 5.81% WER (Extra).”
Acoustically Precise Hesitation Tagging Is Essential for End-to-End Verbatim Transcription Systems, Jhen-Ke Lin et al., arXiv (SLaTE 2025 workshop), 2025, Abstract.
Open the original- What it measures
- In the Pure scheme hesitations are removed. In the Extra scheme they are kept as acoustically precise fillers, which the authors inferred with a large language model from existing audio and transcript pairs.
- What it does not show
- A challenge system on learner speech, not a general benchmark. The difference is small and comes from one corpus.