How to write a custom CutTransform ?

[kobenaxie]

I want to use SoxEffectTransform to perform augmentaions like low_pass or pitch shift in torchaudio, how to write a custom cut_transforms, which can be used like CutMix.

transforms = []
transforms.append(
    CutCustom(...)
)
transforms.append(
    CutMix(...)
)
dataset = K2SpeechRecognitionDataset(
    cut_transforms=ransforms,
)

[pzelasko]

You can check this PR to see the steps needed to add e.g. volume perturbation https://github.com/lhotse-speech/lhotse/pull/382

However, in your case as you're unlikely to modify either sampling rate or num_samples/duration (so the metadata is unaffected), it might be a good idea to implement those as a signal transform instead -- see the ones implemented here https://github.com/lhotse-speech/lhotse/blob/master/lhotse/dataset/signal_transforms.py

They are supported e.g. in K2SpeechRecognitionDataset via input_transforms https://github.com/lhotse-speech/lhotse/blob/master/lhotse/dataset/speech_recognition.py#L65

[kobenaxie]

Thank you for your reply and suggestions, I also find input_transforms in K2SpeechRecoginitionDataset() which is what i need in my case.