Metrics calculation on CPU

[jhauret]

Hi,

I was wondering if someone has implemented a GPU version of torchmetrics.audio.pesq.PerceptualEvaluationSpeechQuality and torchmetrics.audio.stoi.ShortTimeObjectiveIntelligibility which can only be performed on CPU at the moment. If yes, could you make it available for the lightning community ?

Thanks

[SkafteNicki]

Sadly, the PESQ metric is trademarked and therefore cannot be re-implemented on GPU. You can read more here:
https://github.com/ludlows/python-pesq/blob/739759de9dde2d592a6f12b67f3e2875abeecb8f/pesq/dsp.c#L39-L48
STOI could probably be re-implemented in pytorch to gpu accelerate as the implementation seems fairly simple:
https://github.com/mpariente/pystoi/blob/master/pystoi/stoi.py

[jhauret]

I finally used https://github.com/asteroid-team/pytorch_stoi and encapsulated with class inheriting from torchmetrics.metric.Metric