Added WavLM-Base model which replaces the SincNet feature extraction model
within the PyanNet architecture (loaded outside of the class from
HuggingFace.co).

…uding layer selection.

Created a block (in replacement of the old WavLM one) called "selfsup.py" which loads and
apply a specific SSL Torchaudio model, depending on PyanNet's input parameter. User can now
also choose a specific layer which will then be used for feature extraction.
Ex :

seg_model = PyanNet(task=seg, model = "HUBERT_BASE", layer = 5)

This will load "HUBERT_BASE" model and select the 6th layer for the feature extraction. If
layer is not specified, will automatically use the first one (layer 0).

All available models can be found at : https://pytorch.org/audio/main/pipelines.html

… class

Can use pre-trained ssl models from huggingface using PyanHugg class.
Tested (and working) models are :

     - "microsoft/wavlm-base"
     - "microsoft/wavlm-large"
     - "facebook/hubert-base-ls960"
     - "facebook/wav2vec2-base-960h"

Class supports model and layer selection (as well as cache location for the
downloaded model and configuration file).
Ex :
seg_model = PyanHugg(task=seg, selfsupervised={
'model' : 'microsoft/wavlm-base',
'layer' : 2,
'cache' : 'mod_location/'})

Can load a fairseq ckpt from a pretrained model (which is converted to
torchaudio wav2vec2 format)

…fairseq model)

…PyanNetWavLM

…PyanNetWavLM