According to wav2vec model:
it seems that quantitation 
only used on training phase (in order to train the transformer layer to quantitative the data to use small number of discrete values).
So my question is:
when using inference, does this quantitation unit 
being used ? (I guess no, am I right ?)