Negative subtitles duration in SRT

[AlexKhosid]

What happened?

Many subtitle durations in the downloaded SRT are negative and because of it the SRT can't be used.

Steps to reproduce

Run a long transcription (Hebrew) and choose SRT.

What OS are you seeing the problem on?

Window

Relevant log output

App Version: vibe 2.6.6
Commit Hash: 8244a16
Arch: x86_64
Platform: windows
Kernel Version: 10.0.19045
OS: windows
OS Version: 10.0.19045
Cuda Version: n/a
Models: ggml-ivrit-v2-d3-e3.bin, ggml-ivrit-v2-d4.bin, ggml-medium.bin, ivrit-ai--whisper-large-v2-tuned-ggml-model.bin
Default Model: "C:\\Users\\alexk\\AppData\\Local\\github.com.thewh1teagle.vibe\\ggml-ivrit-v2-d4.bin"
Cargo features: vulkan


{
    "avx": {
        "enabled": true,
        "support": true
    },
    "avx2": {
        "enabled": true,
        "support": true
    },
    "f16c": {
        "enabled": true,
        "support": true
    },
    "fma": {
        "enabled": true,
        "support": true
    }
}

[thewh1teagle]

Can you share kink to video / audio that this error happens with?

[thewh1teagle]

I'm pretty sure something wrong with the timestamps in Ivrit model.
I just transcribed 4 hours audio in English and I had no issues saving it in all formats.

[AlexKhosid]

Can you share kink to video / audio that this error happens with?

Yes*: https://mab.to/6XlpYgxWVLPwc/eu1
You can see a negative length of subtitles, for example, in subtitles number 5,7,9,134...
Also, for example, in subtitles 91-99 the length is 0.

*(Please do not use these materials for any purpose other than testing)

[thewh1teagle]

You can see a negative length of subtitles, for example, in subtitles number 5,7,9,134...

I can see in the file but it works fine with other models.

Yes*: https://mab.to/6XlpYgxWVLPwc/eu1 You can see a negative length of subtitles, for example, in subtitles number 5,7,9,134... Also, for example, in subtitles 91-99 the length is 0.

*(Please do not use these materials for any purpose other than testing)

The timestamps looks correct to me but I tried with the default model that comes with Vibe, not with ivrit.ai model.
Maybe the model of ivrit.ai has issues with that. try with this one just for test the timestamps:
https://huggingface.co/ggerganov/whisper.cpp/resolve/main/ggml-tiny-q8_0.bin
Ignore the bad transcription text, focus on the timestamps to see if the 5rd one is correct. to me it looks correct

Also, for example, in subtitles 91-99 the length is 0.

That's issue with whisper.cpp when there's silence. but it shouldn't happen with the default medium model that comes with Vibe

[AlexKhosid]

I tested with several models including the tiny-q8_0.bin you suggested, and found negative timecodes in the generated SRT files with these models as well. I've attached the resulting SRT files for reference:

tests.zip

Tere are negative timecodes for example in subtitles number:
ggml-tiny-q8_0
127,204,290...

ggml-tiny.bin
43,65,125...

ggml-medium.bin
92,146,156...