Linux transcription fails code -6

[jediknight36]

Fresh install of 0.7 on Bazzite. Unzipped the CUDA folder, ran the installer. Loaded a m4a

My specs:

EDIT: using 0.6, it cant find my gpu. Is there a way to point it to it?

[gernophil]

Hey @jediknight36,

what's the output of nvidia-smi? Do you start using the link or using the terminal? What's the output, if you start via Terminal?

[jediknight36]

So I was starting by running the executable from the folder, but I ran it from terminal and that gave me some more info. It said initially that ffmpeg wasnt installed. So I installed it. Still no go. I made sure ffmpeg was installed, python3, pip, and it seems no matter what, I still get the same error. I think it thinks ffmpeg is not installed.

$ ./noScribe

noScribe
AI-powered Audio Transcription 

Fontconfig warning: "/usr/share/fontconfig/conf.avail/05-reset-dirs-sample.conf", line 6: unknown element "reset-dirs"
Hi there, noScribe is ready!

Version 0.7, Kai Dröge 2025
ported to macOS by Philipp Schneider
powered by Whisper (OpenAI), faster-whisper (Guillaume Klein) & pyannote (Hervé Bredin)
https://github.com/kaixxx/noScribe

Audio file(s) selected: 
/var/home/bivey/Downloads/Mental Health Is Important 😊.mp3

Transcript filename(s): 
/var/home/bivey/Downloads/Mental Health Is Important 😊.html

=== Starting queue ===
Processing 1 transcription job(s)

Starting job: Mental Health Is Important 😊.mp3

Converting audio...
Audio conversion finished

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.
          2. The PyTorch version (2.8.0+cu128) 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 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(
pyannote/audio/utils/reproducibility.py:74: ReproducibilityWarning: TensorFloat-32 (TF32) has been disabled as it might lead to reproducibility issues and lower accuracy.
It can be re-enabled by calling
   >>> import torch
   >>> torch.backends.cuda.matmul.allow_tf32 = True
   >>> torch.backends.cudnn.allow_tf32 = True
See https://github.com/pyannote/pyannote-audio/issues/1370 for more details.

  warnings.warn(
pyannote/audio/models/blocks/pooling.py:103: UserWarning: std(): degrees of freedom is <= 0. Correction should be strictly less than the reduction factor (input numel divided by output numel). (Triggered internally at /pytorch/aten/src/ATen/native/ReduceOps.cpp:1839.)
  std = sequences.std(dim=-1, correction=1)
segmentation: 0%segmentation: 51%segmentation: 100%segmentation: 100%segmentation: 100%speaker_counting: 100%embeddings: 0%embeddings: 16%embeddings: 33%embeddings: 50%embeddings: 66%embeddings: 83%embeddings: 100%embeddings: 100%discrete_diarization: 100%

Transcription...
Loading whisper
Voice activity detection...
Transcription...

Unable to load any of {libcudnn_ops.so.9.1.0, libcudnn_ops.so.9.1, libcudnn_ops.so.9, libcudnn_ops.so}
Invalid handle. Cannot load symbol cudnnCreateTensorDescriptor
Transcription worker exited unexpectedly (code -6).
Subprocess terminated unexpectedly
Job error details: Traceback (most recent call last):
  File "noScribe.py", line 3110, in _run_whisper_subprocess_stream
  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 2336, in transcription_worker
  File "noScribe.py", line 2939, in _process_single_job
  File "noScribe.py", line 3123, in _run_whisper_subprocess_stream
Exception: Subprocess terminated unexpectedly


=== Queue processing complete ===
Total jobs: 1
Completed: 0
Failed: 1
Canceled: 0
Total processing time: 0:13
^CTraceback (most recent call last):
  File "noScribe.py", line 3557, in <module>
  File "customtkinter/windows/ctk_tk.py", line 165, in mainloop
  File "tkinter/__init__.py", line 1505, in mainloop
  File "tkinter/__init__.py", line 1963, in __call__
KeyboardInterrupt
[PYI-45914:ERROR] Failed to execute script 'noScribe' due to unhandled exception!
multiprocessing/resource_tracker.py:279: UserWarning: resource_tracker: There appear to be 1 leaked semaphore objects to clean up at shutdown
`

Im running this from the folder. 

So I decided to run a distrobox, basically a vm, of Ubuntu. Its all in terminal, so I installed manually from source code in ubuntu.  There, Im told to run cd into noScribe and run "python3 ./noScribe.py" but when I do, it always says Im missing modules. Ive been trying to install them as I go, but it seems like a fools errand. The latest error is this:
`$ python3 ./noScribe.py
Traceback (most recent call last):
  File "/var/home/bivey/noScribe/./noScribe.py", line 34, in <module>
    from PIL import Image
ImportError: cannot import name 'Image' from 'PIL' (unknown location)

Im lost at this point. I even tried copying everything over from the noScribe I downloaded and ran from originally.

when I run nvidia-smi inside the container, I get this:

When I run it in Bazzite, I get:

Im not sure why it says the gpu is "off" either. I can game with modern games just fine.

[gernophil]

You can ignore the ffmpeg warning. This is just a warning from the package torchcodec, but we use an internal ffmpeg for 0.7.0 and since 0.7.1 we don't use ffmpeg at all anymore.

The thing that seems to be missing is cuDNN (and maybe the CUDA toolkit). I would start by installing those.

[gernophil]

@kaixxx, when I checked yesterday the CUDA version of 0.7.1 wasn't available. Could you upload it?

[gernophil]

@jediknight36: running it from source via python need a little more then just installing python. You also need to create a venv with the necessary modules/packages. A tutorial should be somewhere in the readme.

[kaixxx]

noScribe relies on some libraries for CUDA support that must be installed on the system. @gernophil has already found out that the library cuDNN 9 for CUDA 12 is missing on your system. Normally, I would point you to https://docs.nvidia.com/deeplearning/cudnn/installation/latest/linux.html to install it.
However, your case is a little more complicated, because Bazzite seems to be an "immutable" image based OS, where you cannot add your own system libraries. Here is a possible workaround:

# 1. Make a small venv just to host the CUDA libs
python3 -m venv ~/noscribe-cuda-libs
source ~/noscribe-cuda-libs/bin/activate

# 2. Install the runtime libraries Faster-Whisper expects
pip install -U pip
pip install nvidia-cublas-cu12 "nvidia-cudnn-cu12==9.*"

# 3. Point the dynamic loader at those libraries
export LD_LIBRARY_PATH="$(python3 -c 'import os, nvidia.cublas.lib, nvidia.cudnn.lib; print(os.path.dirname(nvidia.cublas.lib.__file__)+\":\"+os.path.dirname(nvidia.cudnn.lib.__file__))')${LD_LIBRARY_PATH:+:$LD_LIBRARY_PATH}"

# 4. Start noScribe from the same shell
cd /path/to/noScribe
./noScribe

To verify that cuDNN is actually present before starting noScribe:

python3 -c "import os, nvidia.cudnn.lib; print(os.path.dirname(nvidia.cudnn.lib.__file__))"
find "$(python3 -c 'import os, nvidia.cudnn.lib; print(os.path.dirname(nvidia.cudnn.lib.__file__))')" -name 'libcudnn_ops.so*'

If that works, make it persistent by putting the source ~/noscribe-cuda-libs/bin/activate and export LD_LIBRARY_PATH=... lines into a small wrapper script.

The only other option I see would be to run the transcription on CPU only, which is mmuch slower. You can do that by editing the file ~/noScribe/noScribe/config.yml and set the value force_whisper_cpu to 'True' (keep the quotes). Make sure to edit this file only when noScribe is closed.