Piotr Żelasko
Piotr Żelasko
Would you be willing to contribute the fix? You'd need to modify the method here: https://github.com/lhotse-speech/lhotse/blob/e982003c1777a8f48ee40580bd835c931a03062d/lhotse/dataset/signal_transforms.py#L262
You can move the Wav2Win (or any other module in that file) to device before running the inference. If you're using `Fbank`, there's an option in the config to place...
Wouldn't `module = module.to("cuda")` solve the issue by placing every registered parameter and buffer on the GPU?
I don't think `decompose` was ever tested in this way, although I would have expected it to work. I'm afraid I don't have enough time right now to look into...
Thanks, you're right. I'll keep the issue open for now.
Features does have ` recording_id` field. If you can provide some way to reproduce with a small dataset like yesno or mini Librispeech I can look into it.
You can use DynamicCutSampler and specify max_cuts rather than max_duration.
That's surprising, I don't think there's any code in lhotse itself that would change the encoding. If the env var works for you that sounds good.
Update the paths to the h5 files in your manifests.