Is torchaudio 2.9 supported?

[submagr]

Tested versions

• Reproducible in 4.0.3

System information

Linux 5.15.0-76-generic #83-Ubuntu SMP Thu Jun 15 19:16:32 UTC 2023 x86_64 x86_64 x86_64 GNU/Linux

Issue description

I notice that pyannote explicitly depends on torchaudio 2.8. Is it possible to use pyannote with torchaudio 2.9?

Details:

• A lot of torchaudio's functionalities were dropped in version 2.9
• torchaudio 2.8 does not work with torch 2.9
• My other libraries depends on torch 2.9

Minimal reproduction example (MRE)

This is a general question

[mobabel]

Same question here, I have the Torch 2.9.1+cuda129, for using payannote-audio I have to downgrade my torch, it is sad.

[submagr]

@mobabel I was able to get pyannote-audio working with torch 2.9. I tested it for diarization task:

>>> torch.__version__
'2.9.1+1765427380.cuda130.cudnn916.ap'
>>> torchvision.__version__
'0.24.0+7a9db90'
>>> torchaudio.__version__
'2.9.0+eaa9e4e'

---

• These instructions will install torchcodec==0.9.0 which might create issues with other packages installed in your
  local environment. I recommend doing everything in a fresh virtual environment as following:

python3 -m venv venv_pyannote_audio
source venv_pyannote_audio/bin/activate

• Install torch 2.9 and torchaudio 2.9

• Install pyannote.audio v4.0.3 without dependencies (we will manually install required dependencies):
  pip install --no-deps pyannote-audio==4.0.3

• Install dependencies:

  • Note: pyannote-audio uses torchcodec==0.7.0 but we need 0.9.0 for working with torchaudio 2.9
  • Note: Use my speechbrain fork that fix issues with torchaudio 2.9

pip install --no-input \
    asteroid-filterbanks>=0.4.0 \
    opentelemetry-api>=1.34.0 \
    opentelemetry-sdk>=1.34.0 \
    opentelemetry-exporter-otlp>=1.34.0 \
    pyannote-core>=6.0.1 \
    pyannote-database>=6.1.1 \
    pyannote-metrics>=4.0.0 \
    pyannote-pipeline>=4.0.0 \
    torch-audiomentations>=0.12.0 \
    torchcodec==0.9.0 \
    speechbrain @ git+https://github.com/submagr/speechbrain.git@252bb77f05916b1979115691b837824e98ec40b4

[mobabel]

@submagr Thank you for your feedback, I have tried your solution, but got this errors:

ERROR: pip's dependency resolver does not currently take into account all the packages that are installed. This behaviour is the source of the following dependency conflicts. pyannote-audio 4.0.3 requires einops>=0.8.1, which is not installed. pyannote-audio 4.0.3 requires huggingface-hub>=0.28.1, which is not installed. pyannote-audio 4.0.3 requires lightning>=2.4, which is not installed. pyannote-audio 4.0.3 requires matplotlib>=3.10.0, which is not installed. pyannote-audio 4.0.3 requires pyannoteai-sdk>=0.3.0, which is not installed. pyannote-audio 4.0.3 requires pytorch-metric-learning>=2.8.1, which is not installed. pyannote-audio 4.0.3 requires rich>=13.9.4, which is not installed. pyannote-audio 4.0.3 requires safetensors>=0.5.2, which is not installed. pyannote-audio 4.0.3 requires soundfile>=0.13.1, which is not installed. pyannote-audio 4.0.3 requires torchmetrics>=1.6.1, which is not installed. pyannote-audio 4.0.3 requires torch==2.8.0, but you have torch 2.10.0+cu129 which is incompatible. pyannote-audio 4.0.3 requires torchaudio==2.8.0, but you have torchaudio 2.10.0+cu129 which is incompatible. pyannote-audio 4.0.3 requires torchcodec==0.7.0, but you have torchcodec 0.9.0 which is incompatible.