The time precision in generated subtitle files is second, not millisecond

[whiteeat]

As the title says, the time precision in generated subtitle files is not milliseconds. So my question is, whether the code just work like that, or it is a bug? And, if I want the precision to be milliseconds, how can I achieve that?

[glangford]

From a comment by @jongwook in another discussion,

Currently, the predicted timestamps tend to be biased towards integers, and there are some failure modes where the timestamps can be constantly shifted

See the discussion "Improving Timestamp Accuracy" in #435 where this is covered in detail.

For word level timestamps see #26 and this recently added notebook.
https://github.com/openai/whisper/blob/28769fcfe50755a817ab922a7bc83483159600a9/notebooks/Multilingual_ASR.ipynb

[whiteeat]

Thanks very much! I try solutions above and stable-ts works best for me!

[kevinsaracho]

Como puedo solucionar el problema?

[glangford]

@kevinsaracho Use --word_timestamps True