CPC label about s3prl

[andi611]

Hello, Author ! Thank you for your sharing !

I have some problems about the CPC label. I find that code in 'dataloader.py'.

assert('train-clean-100' in sets and len(sets) == 1)

Does this code mean that I can only use the 'train-clean-100' ?

Yes, because in the CPC paper Section 3.1 it says:
"For audio, we use a 100-hour subset of the publicly available LibriSpeech dataset [30]. Although the
dataset does not provide labels other than the raw text, we obtained force-aligned phone sequences with the Kaldi toolkit [31] and pre-trained models on Librispeech. We have made the aligned phone labels and our train/test split available for download on Google Drive."

Hence, the train-clean-100 is split into train/test sets. Here we follow the setting of DeepMind.

In addition, I have use kaldi to align the LibriSpeech but I can not get the file 'converted_aligned_phone.txt'. How can I obtain the CPC label of other datasets ?

The file converted_aligned_phone.zip is provided in this repo under data/cpc_phone. See this section of README.

I hope this helps!

[developerFanYu]

Thank for your answer!

However, I also have another question: I would like to use 'train-clean-360' or ''train-other-500 to start my experiments. How can I obtain the 'converted_aligned_phone.txt' of these datasets.

[andi611]

Thank for your answer!

However, I also have another question: I would like to use 'train-clean-360' or ''train-other-500 to start my experiments. How can I obtain the 'converted_aligned_phone.txt' of these datasets.

The file converted_aligned_phone.txt is generated with Kaldi, and provided by DeepMind in the CPC paper. I simply downloaded the file from DeepMind's google drive, I do not have experience with this.