Preprint
Whisper Has an Internal Word Aligner
Sung-Lin Yeh et al., arXiv (ASRU 2025), 2025
Read the originalRead on October 17, 2026
- What it is
- A 2025 study of word-level timestamps from Whisper. It finds attention heads that give accurate word alignments without training, and it compares its error against a stricter tolerance than earlier studies used.
- How much weight it carries
- Peer-reviewed conference version (ASRU 2025) of a preprint. It tests on its own evaluation setup, so its error figures are comparable to other papers only when the tolerance is the same.
Facts taken from this source
- Earlier studies of word-level timestamps usually counted a timestamp as correct if it fell within a tolerance of more than 200 milliseconds.
“The evaluation in prior work is also relatively loose, typically using a tolerance of more than 200 ms.” (Abstract)
- Under a stricter tolerance of 20 to 100 milliseconds, the 2025 study's word alignments from Whisper were more accurate than those of earlier work.
“produces word alignments that are more accurate than prior work under a stricter tolerance between 20 ms and 100 ms” (Abstract)