can i preprocess the libri_mel160_subword5000 with the preprocess_libri script? about s3prl

[ArtemisZGL]

Beside, I also want to know about the meaning of name of pretrain ckpt, such as M9 and M6. Thanks.

[andi611]

I have the librispeech dataset, but when i download the feature from the google drive, it's slow. So I wonder can I process it with the original dataset?

If you want to use features other then fMLLR, yes you can use the preprocess/preprocess_libri.py. See the instructions in README.md. In particular, this line will preprocess the ibri_mel160_subword5000 data:
python preprocess/preprocess_libri.py --feature_type=mel --data_path=../data/LibriSpeech

If you want to use fMLLR, please refer to this issue. In short, you have to download the fMLLR from google drive if you want to use our pre-trained models.

I'm a little confuse about the "subword" meaning for mel feature.

Please ignore this, it is just a naming issue in our early development stage. The transcriptions are also preprocessed in subword units, but they were never used and deprecated.

Beside, I also want to know about the meaning of name of pretrain ckpt, such as M9 and M6. Thanks.

M9 means masking 9 frames, M6 for 6 frames.

Hope this helps.
Andy

[ArtemisZGL]

Thank you for the detailed explanation! It's really a great job!