Preprint
Acoustically Precise Hesitation Tagging Is Essential for End-to-End Verbatim Transcription Systems
Jhen-Ke Lin et al., arXiv (SLaTE 2025 workshop), 2025
Read the originalRead on October 17, 2026
- What it is
- A 2025 study that fine-tunes Whisper models for verbatim transcription of non-native English speech, comparing three ways of labeling hesitations and fillers.
- How much weight it carries
- Peer-reviewed workshop paper on the Speak and Improve 2025 corpus. Results are for non-native learner speech and a specific fine-tuning setup, so they do not generalize to every recording.
Facts taken from this source
- 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.
“Our challenge system achieved 6.47% WER (Pure) and 5.81% WER (Extra).” (Abstract)
- Many ASR systems discard or generalize hesitations, which loses acoustic detail that verbatim transcription depends on.
“many ASR systems discard or generalize hesitations, losing important acoustic details” (Abstract)
- Fine-tuning Whisper Large V3 Turbo on labeled filled pauses gave 5.5 percent word error rate, against 6.2 percent when hesitations were removed.
“fine-tuning Whisper Large V3 Turbo with the "Extra" scheme yielded a 5.5% WER, an 11.3% relative improvement over the "Pure" scheme (6.2% WER).” (Abstract)