avoid audio clipping of generated audio

[defgsus]

Hi and thanks! This model is fun! On the :hugs: space, some generated samples have very strong clipping. While other samples are very quiet.

Could you add a volume normalization before converting it to wav?