predict the accurate end timestamp of a word

[climber418]

If there is a slightly long silence gap between two words, how can statle-ts get the accurate end timestamp of a word?

[sockthem]

facing same issue, predicted timestamps are not reliable.

[tohe91]

I have the same issue with longer silences, I tried using vad and slicing the audio file to then transcribe the slices and use the vad timestamps, but the transcription quality got very bad even with using prompt. I'm facing mainly too problems, first being the described issue with longer silences and second that sometimes words from a duration of 10-15 seconds all get the same timestamp.

[jianfch]

Have you tried the default settings in stable-ts?
If you're still getting this issue, you can change settings to make it more aggressive.

https://github.com/jianfch/stable-ts/blob/main/stable_whisper/whisper_word_level.py#L124-L137

from stable_whisper import load_model
model = load_model('base')
results = model.transcribe('audio.mp3',  remove_background=True, silence_threshold=1.0)

If the gaps are not completely silent you can also try to increase the lower_threshold.
You can also try it with different size models.
Also try to avoid using beam search (beam_size) because it makes the gap suppression less effective.