Slower performance for Torchmetrics PESQ

[ashinkajay]

🐛 Bug

Multiprocessing capability is missing for Torchmetrics PESQ while the underlying library ludlows/python-pesq have that option.
As a result, the torchmetrics version of PESQ performs slower than ludlows/python-pesq.

To Reproduce

Run PESQ for a batch of 100 audios with a minimum duration of 10 seconds for each audio.

Code sample

Python notebook code attachment here:
analysis.zip

Expected behavior

Both the Torchmetrics PESQ and ludlows/python-pesq should take almost the same time to compute the PESQ score.

Environment

• Python - 3.10.4 (installed using conda)
• OS: Ubuntu 20.04.4 LTS

Below libraries installed using pip

• TorchMetrics - 0.9.3
• PyTorch - 1.12.1
• Torchaudio - 0.12.1
• PESQ - 0.0.4

Additional context

References
ludlows/python-pesq: https://github.com/ludlows/PESQ

[stancld]

Hi @ashinkajay, thank you for your post. As far as I know, the original PESQ code has a pretty restricting license regarding the way how PESQ can be implemented in torchmetrics (see discussion in #726)

cc: @Borda @SkafteNicki Have you had any discussion about this besides the issue attached? :]

[SkafteNicki]

Yes their license is very strict so we cannot re-implement. However, it seems like we could probably call pesq_batch from https://github.com/ludlows/PESQ instead of pesq to run the calculation in parallel. We would just need to add an n_processor argument to the modular and functional implementations.

[SkafteNicki]

Yes their license is very strict so we cannot re-implement. However, it seems like we could probably call pesq_batch from https://github.com/ludlows/PESQ instead of pesq to run the calculation in parallel. We would just need to add an n_processor argument to the modular and functional implementations.

@ashinkajay is this what you are thinking about when raising this issue?

[ashinkajay]

Thank you @SkafteNicki and @stancld

@SkafteNicki Yes I found pesq_batch and couldn't see that functionality in torchmetrics. That is why I raised the issue.
Thanks for the clarification about the license.