Linux Arch: Transcription fails with 0.7

[kmille]

Hey,
I'm the Arch maintainer and wanted to update to 0.7. This the error I get:

[kmille]

Speaker identification...
Loading pyannote
Fontconfig warning: "/usr/share/fontconfig/conf.avail/05-reset-dirs-sample.conf", line 6: unknown element "reset-dirs"
pyannote/audio/core/io.py:47: UserWarning: 
torchcodec is not installed correctly so built-in audio decoding will fail. Solutions are:
* use audio preloaded in-memory as a {'waveform': (channel, time) torch.Tensor, 'sample_rate': int} dictionary;
* fix torchcodec installation. Error message was:

Could not load libtorchcodec. Likely causes:
          1. FFmpeg is not properly installed in your environment. We support
             versions 4, 5, 6, and 7 on all platforms, and 8 on Mac and Linux.
          2. The PyTorch version (2.8.0+cpu) is not compatible with
             this version of TorchCodec. Refer to the version compatibility
             table:
             https://github.com/pytorch/torchcodec?tab=readme-ov-file#installing-torchcodec.
          3. Another runtime dependency; see exceptions below.
        The following exceptions were raised as we tried to load libtorchcodec:
        
[start of libtorchcodec loading traceback]
FFmpeg version 8: No spec found for libtorchcodec_core8
FFmpeg version 7: No spec found for libtorchcodec_core7
FFmpeg version 6: No spec found for libtorchcodec_core6
FFmpeg version 5: No spec found for libtorchcodec_core5
FFmpeg version 4: No spec found for libtorchcodec_core4
[end of libtorchcodec loading traceback].
  warnings.warn(
torchaudio/_backend/utils.py:213: UserWarning: In 2.9, this function's implementation will be changed to use torchaudio.load_with_torchcodec` under the hood. Some parameters like ``normalize``, ``format``, ``buffer_size``, and ``backend`` will be ignored. We recommend that you port your code to rely directly on TorchCodec's decoder instead: https://docs.pytorch.org/torchcodec/stable/generated/torchcodec.decoders.AudioDecoder.html#torchcodec.decoders.AudioDecoder.
  warnings.warn(
Diarization worker exited unexpectedly (code -11). UI remains responsive.
Error in step 2 - identifiying the speakers.: Subprocess terminated unexpectedly
Job error details: Traceback (most recent call last):
  File "noScribe.py", line 3213, in _run_diarize_subprocess
  File "multiprocessing/queues.py", line 114, in get
_queue.Empty

During handling of the above exception, another exception occurred:

Traceback (most recent call last):
  File "noScribe.py", line 2624, in _process_single_job
  File "noScribe.py", line 3224, in _run_diarize_subprocess
Exception: Subprocess terminated unexpectedly

During handling of the above exception, another exception occurred:

Traceback (most recent call last):
  File "noScribe.py", line 2336, in transcription_worker
  File "noScribe.py", line 2648, in _process_single_job
Exception: Error in step 2 - identifiying the speakers.: Subprocess terminated unexpectedly

Can you tell me the dependecies? Seems like something is missing. v0.6.2 just worked.

[gernophil]

Yes, that's a similar error I get with Fedora. Don't know why this happens yet. For some users it works just fine and for others not at all as you can see here.

Are you using the 0.7beta or final? And the CPU or the CUDA version. There was a bug with the logo.png file in some of the installers, that's why I'm reuploading the CUDA version right now (CPU is reuploaded already).

You can try to run it from source with these requirements.

[gernophil]

I think it might be related to this, but I can't put my finger on it.

[kmille]

I'm using the CPU version and noScribe_0.7.0_cpu_linux_amd64.tar.gz (6d20512ff316dc937f3ccd4c83779197b73e0dc35f5282b702823a76059cad28).

[gernophil]

The venv I used to create the bundles use

torch==2.8.0 # or torch==2.8.0+cpu
torchcodec==0.8.1

This table indicates that torch 2.8 is compatible with torchcodec 0.6 and 0.7, but not 0.8, but on the other hand on macOS this combination is working and torchcodec 0.8.1 was installed as direct implicit dependency by pip, so it should (hopefully) work.

Could you try installing ffmpeg with your package manager (e.g. pacman) and see, if this solves the issue?

[gernophil]

Also found this along the way: meta-pytorch/torchcodec#1108

[gernophil]

So, it seems this combination might be part of the issue: meta-pytorch/torchcodec#1110 (comment)

@kaixxx, in this case I might need to rebuild all macOS and Linux bundles. In your windows build what combinations do you use? Also IIRC pyannote4 was not compatible to torch 2.9. Has this changed?

also @kaixxx, the error message states as possible solution: * use audio preloaded in-memory as a {'waveform': (channel, time) torch.Tensor, 'sample_rate': int} dictionary;. Isn't that, what you are doing?

[kaixxx]

I have tried to work around torchcodec. Instead, I do what the error message suggests: preload the audio into memory first and send this waveform to pyannote, not the filename:

waveform, sample_rate = torchaudio.load(audio_file)
...
diarization = pipeline({"waveform": waveform, "sample_rate": sample_rate}, hook=hook, num_speakers=num_speakers)

(see pyannote_mp_worker.py, line 81ff.)

In my env on Windows, torchcodec 0.7.0 was still installed for some reason. I have now removed it completely and noScribe is still running perfectly. So, it is really not used.

The reason to work around torchcodec was that it depends on a system-wide ffmpeg install (at least this is how I understand it). I wanted to avoid such an external dependency.

[gernophil]

Maybe this workaround somehow doesn't work in some Linux and x86_64-macOS, because I'm getting the same error there. It seems like at some point torchcodec is unintentionally loaded.

[kaixxx]

The complaint about torchcodec is just a warning, this should not crash the whole subprocess. Problem is that we have so little information, what else might be the reason. The return code -11 points to a SEGFAULT, but that is not telling us much.

@kmille: Could you try running a transcription with speaker detection set to "none"? This would skip the whole pyannote process and tell us if at least whisper is working (which I doubt...).

[kmille]

Everything works fine when I set speaker detection to "none".

[gernophil]

Just wanted to point out that I falsely assumed this error is similar to the one I get with my Fedora VM. For me speaker detection works, but transcription fails :D.

[kaixxx]

Everything works fine when I set speaker detection to "none".

Thank you, that tells us that there is nothing fundamentally wrong PyTorch as this is also used by faster-whisper.

I think I have found the root cause:

While looking around the pyannote issues, I came across this, which fits very well our problem discussed here - the incompatibility between version 0.8 of torchcodec and torch version 2.8.0, leading to a segmentation fault: meta-pytorch/torchcodec#995 So, the solution might be to downgrade torchcodec to 0.7.0 or to remove it completely from the build. Your idea above was right, @gernophil (#256 (comment))!

[gernophil]

I'll quickly recompile a linux CPU with torchcodec 0.7. Let's see, if this helps.