search

haiciyang / Feature-predictor-for-speech-codec

MIT License
7 stars 0 forks source link

What are the pre-computed features in 'features.pt'? #6

Open rppravin opened 4 months ago

rppravin commented 4 months ago

Thanks for the code.

In dataset_orig.py, looks like some pre-computed features are assumed to be stored in 'features.pt'. Can you please let me know what these features are? Are these Cepstral features computed from Librispeech data using a different code?

Thanks in advance!

haiciyang commented 3 months ago

Hi! Yes the features are cepstral features of the speech signal from the training dataset. They were exacted the same way as LPCNet does. LPCNet's repo (https://github.com/xiph/LPCNet) has provided the code. You can refer to the second step of Training a new model - generate training data ((https://github.com/xiph/LPCNet))

rppravin commented 3 months ago

Thank you! Can you also confirm the input features: [0 to 17] - Bark Cepstrum, [18] - pitch lag, [19] - pitch correlation. Is this correct?

haiciyang commented 3 months ago

Yes. That's correct. The same with LPCNet's original feature design.