when I transcribe japanese video, sometime whole script file repeat same dialoge

[kiron111]

for example:
1
00:00:00,000 --> 00:00:02,000
I'm not sure if I'm going to be able to get through this.

2
00:00:02,000 --> 00:00:04,000
I'm not sure if I'm going to be able to get through this.

3
00:00:04,000 --> 00:00:06,000
I'm not sure if I'm going to be able to get through this.

4
00:00:06,000 --> 00:00:08,000
I'm not sure if I'm going to be able to get through this.

5
00:00:08,000 --> 00:00:10,000
I'm not sure if I'm going to be able to get through this.

6
00:00:10,000 --> 00:00:12,000
I'm not sure if I'm going to be able to get through this.

Sometime, it happen at the beginning, sometime at the end (maybe last for 10 minutes), sometime the whole file repeat same dialogue. (no matter I use medium or large model)

When I use the whisperer, using the same model, everything is okay.

So, I think maybe the optimization is over, skipping some important step.

[tigros]

that's probably the --max-context 0 parameter which i saw somewhere and added it to whisperer.

[kiron111]

I'm looking forwards for the next update 👍

[emcodem]

#26

[Taounit]

After making some tests I am pretty confident the problem comes from the translation module (Google API ?) and not whisper, as the repeat patterns occur in translation mode and stop when the transcription-only mode is used.

[albino1]

Unfortunately, this happens all the time without any translation at all. Check this github, the main whisper.cpp one, and even the official OpenAI one, there's probably a hundred posts about it across them.

[jiyuguan]

I’m writing this comment to report a problem with the whisper app. I found that when it transcribes voice to text, it repeats the same sentence until the end of the video. It only transcribes the first part of the voice accurately. I tried to cut the part that was not transcribed properly, and it worked. But then the same problem happened again in the second half of the edited video. I hope you can fix this issue soon.

[Highlander1536]

It can get really bad with live audio translating, anyone know a way to lower the repeating at least?

[emcodem]

@Highlander1536 sure, i have limited it with a dirty hack: #26 - what i do is detect if repeated text has been decoded and if yes, reset the context history.
Also, to repeat above stuff: -mc 0 seems to always solve it (but lowers quality, maybe --prompt can help here?)