Word level timestamp

[RobinWitch]

I‘m trying to use it to get each word's timestamp, but my result is bad. I'm not sure what I did wrong.
I use whisper-large-v3
openai-whisper 20231117
whisper-timestamped 1.15.4

I use the command whisper_timestamped test.wav --language en --model ./ckpt/whisper-large-v3 --output_dir output1 --accurate
Wav file and the generated result are placed in wt_test.zip.

For example. in test.wav.words.tsv, the text relaxing starts at 3420, ends at 3800. But actually, the time range for the pronunciation of 'relaxing' should be between 3600 and 4300. The error is so big.

Can someone tell me how to improve it and get a good result?

[Jeronymous]

Can you try with option --vad True ?

https://github.com/linto-ai/whisper-timestamped?tab=readme-ov-file#options-that-may-improve-results