Whisper fails on some files, but works fine on others

[Cube-Oakley]

Describe the bug
When searching for subtitles in Bazarr and having the Whisper provider set up, Whisper will repeatedly fail to provide the subtitles for specific files and Bazarr will then throttle Whisper as a subtitle provider

To Reproduce
Steps to reproduce the behavior:

1. Configure Whisper as a provider
2. Click on search on an affected TV Show or Movie (This only affects specific files, other files work fine)
3. After 30 seconds to a minute the Whisper provider will be throttled, and the subtitle is not generated

Expected behavior
When clicking on search for subtitles for a specific movie or show, it should communicate with Whisper properly and generate subtitles

Software (please complete the following information):

• Bazarr: v1.4.2
• Radarr version: 4.5.2.7388
• Sonarr version: 3.0.10.1567
• OS: Docker hosted on Ubuntu Server w/Portainer

Additional context
This is my first ever bug report so I apologize if there's anything I've done incorrectly. I read through all related issues and I found one semi-similar, but it was closed without a fix as the original submitter never responded. Also searched Reddit/Internet and did a fair amount of troubleshooting with no fix.

• Have tried updating Bazarr (was using an older version)
• Deleted the database db file and allowed Bazarr to recreate to rule out database corruption
• Specific Whisper provider in use is "Subgen" also running in Docker
• The majority of files transcribe fine. If I go to a random file and try and transcribe it, there's a high chance it'll work. But when I hit "Search All" I've had times where it will pick a specific file to transcribe first that fails. It then will continually fail unless I search for subtitles episode by episode, or if I remove the language profile from that specific bad episode allowing the "search all" to bypass the one it fails on
• I've verified that the files themselves are fine. They play fine, and aren't in any different format than the files that work
• Very repeatable. If a file fails, it will always fail no matter how many times I re-try the operation
• I've verified that there is English audio within the first 30 seconds of the video
  BazarrLogs.txt

[McCloudS]

Your language in the file shows as 'an' or Aragonese (from FFProbe), which is not supported by Whisper, which is why you're getting an error. I imagine that's wrong, but it's the source of truth feeding into the Whisper Provider. Probably needs to be an update on the Whisper Provider side to gracefully handle it.

@ayancey

[Cube-Oakley]

Ah I see, that makes sense. I can confirm that the file is indeed English and not Aragonese, so Whisper must be recognizing it wrong.

I've done a little further troubleshooting related to the settings of the Whisper container including changing the detection interval up to 60 seconds from the default 30, and attempting to force detection as English.

After making these changes I'm still unable to generate subtitles for the particular file in the logs I attached, but I was able to generate them for a few other files that failed prior. So it seemed to help, but not entirely fix the issue.

I'm not positive but I think when Bazarr passes the data to Whisper to transcribe, Whisper successfully takes into account the increased detection time, but ignores the forced language setting (this particular transcription still fails due to mis-recognizing the language).

Subgen (this particular whisper provider) does have the option to transcribe directly from a folder. I tested pointing it to my TV library and it looks like the forced language setting did work then, but there wasn't any way for me to force transcription of this specific file and it sort of started transcribing whatever it felt like. When I have a moment I'll see if I can setup a dedicated docker vol with this one particular file in it to see if it can transcribe directly with and without the forced language setting enabled.

Two additional things I've tried since the original report are:

1. Updating Subgen container to latest version (no difference)

2. Setting up another docker container with the official Whisper-asr-webservice and setting it as the Whisper provider in Bazarr (same exact Aragonese mis-detection issue, so the issue isn't Subgen specific)

One last thing to note, in all of these tests, the model in use is the "medium - FasterWhisper" model. Have not tried other models yet.

Thanks!

Editing to say that when I wrote the above, I didn't realize you were the dev of Subgen! That's really cool, and I love the app!

[McCloudS]

It isn’t a mis-detection issue from Whisper. Your file has the audio stream set as AN. You can find a program to modify the file to change it back to the appropriate language or download a new file that has the audio set correctly.

Even if the Bazarr Whisper Provider is updated to gracefully handle your error, your file still won't be able to have a subtitle generated because the FILE has the wrong language set.

[hnorgaar]

This subject have been up several times and there is no way to force Bazarr to use anything else than detection. There is a way to specify/change audio language in Sonarr which Bazarr see, however not take into account. Maybe make a request for that to be recognized and prioritized so its possible to gibe Bazarr a pointer

[Cube-Oakley]

This makes sense, and upon more careful inspection with ffprobe I can confirm that this particular file does have "an" named as the language even though it's English when played.

I'll see if I can change it to "en", and if I can't I'll just redownload a working file. I think we are good to close this as it's essentially a broken file. The other files I was originally having issues with were able to be transcribed after changing the detection interval to 60 seconds.

I'll consider putting in a feature request for a forced audio language in Bazarr if this is something that ends up coming up a lot.

Thanks!