Timestamps Accuracy

[ArieBen]

Subtitles are often timestamped with durations that start significantly earlier than when the words are actually spoken. The duration of subtitles can be prolonged in two ways: either with a single long timespan or through a series of chained identical subtitles.

Programmatically trimming the beginning or end of these timespans or chains often produces unexpected results.

Compared to the original OpenAI implementation, hallucinations are significantly reduced with Whisper.cpp. Additionally, I can utilize larger models, but the timestamp accuracy is still lacking.

Do you have any suggestions for improving this?