Fact · Transcription
In a 2024 Interspeech study, adjusting the tokenizer of the Whisper model significantly improved the precision of its word-level timestamps.
Checked against its source on October 17, 2026
In the source's own words
“We demonstrate that carefully adjusting the tokenizer of the Whisper speech recognition model significantly improves the precision of word-level timestamps”
CrisperWhisper: Accurate Timestamps on Verbatim Speech Transcriptions, Laurin Wagner et al., Interspeech 2024, 2024, Abstract.
Open the original- What it measures
- Word timestamps from Whisper are not built in. They come from a second step that reads the model's attention, and the tokenizer affects how well that step works.
- What it does not show
- The gain is measured on a fine-tuned, more verbatim model, not on the stock Whisper release.