"Субтитры подогнал «Симон»" - dirty datasets for Russian subtitles.

[LocalVoidPictures]

On multiple occasions, I have received "nonsensical" results, such as the one in the title (read literally - "Subtitles by Simon"). You can frequently find this line by doing a simple online search. This phrase has nothing in common with actual transcription. Including it in the dataset results in wildly incorrect results.

Funnily enough, I'm only getting this when using a large model, not medium or tiny.

Where do we report this?

Here's a sample of what I'm getting sometimes:

% whisper "video.mov" --model large --language Russian
/Library/Frameworks/Python.framework/Versions/3.12/lib/python3.12/site-packages/whisper/transcribe.py:115: UserWarning: FP16 is not supported on CPU; using FP32 instead
  warnings.warn("FP16 is not supported on CPU; using FP32 instead")
[00:00.000 --> 00:29.980]  Субтитры подогнал «Симон»
[00:30.000 --> 00:59.980]  Субтитры подогнал «Симон»
[01:00.000 --> 01:29.980]  Субтитры подогнал «Симон»
[01:30.000 --> 01:59.980]  Субтитры подогнал «Симон»
[02:00.000 --> 02:29.980]  Субтитры подогнал «Симон»
[02:30.000 --> 02:59.980]  Субтитры подогнал «Симон»
[03:00.000 --> 03:29.980]  Субтитры подогнал «Симон»
[03:30.000 --> 03:59.980]  Субтитры подогнал «Симон»
[04:00.000 --> 04:29.980]  Субтитры подогнал «Симон»
[04:30.000 --> 04:59.980]  Субтитры подогнал «Симон»
[05:00.000 --> 05:29.980]  Субтитры подогнал «Симон»
[05:30.000 --> 05:59.980]  Субтитры подогнал «Симон»
[06:00.000 --> 06:29.980]  Субтитры подогнал «Симон»
[06:30.000 --> 06:59.980]  Субтитры подогнал «Симон»

[itaipee]

large-v2 has much lesser hallucinations than the original large model

please try
whisper "video.mov" --model large-v2 --language Russian

if it does not help , you need to research a bit on hallucinations , there are several posts on the subject.

p.s. do the 7 minutes in the video include any speech , or just silence or music ?

[LocalVoidPictures]

In fact I've used large-v3, which is what is automatically downloaded by the CLI tool whenever I set --model large. Perhaps though it's not as straightforward as it sounds? I'll try your suggestion nevertheless.

The audio is chock-full of text, except for the initial 2 minutes (ambience). However, smaller models pick up on it correctly (they just make too many mistakes in the words themselves, which is why I attempted a larger model).

I'll check out the posts on hallucinations of course. However, I believe that the issue I raised is important regardless. It's not a random hallucination, it's obviously a part of a larger issue. You can find this exact phrase all over the internet with a simple search. For example, check here (Ctrl+F for Субтитры подогнал «Симон»). Obviously it introduces unnecessary noise into the model. The question is whether this is the correct place to raise this topic.