After lots of trial and error I gave up on using the standard tensorflow framework, It didn't seem to matter which optimizer, which learning rate, how many epochs, I just could not generate anything usable. I spent quite a bit of time on the mel-specs and reviewing what I could, the other issue here was the visibility of the data before the training. This made it time consuming to interate over.

Regigging the notebooks

I became paranoid about data quality and wanted to make sure that the data that had been sorted into the training folders was indeed correct. Data visualizatiion before training became pretty important to me. I spent some time making sure I could review samples before training started and now I get a consistent output to review.

Mel-specs output:

Fastai and fastaudio early training

Fastai and fastaudio try to abstract alot of the detailed requirements from the training so that you start with a good base level. You can still tweak it but you are less likely to end up with results like i did with tensorflow.

Needs more tweaking, but in general a much more positive expierence and higher level of confidence that we are on the right track.

The other key thing to note here, we no longer generate mel-specs or need to do preprocessing. This is done on the fly. Fastai and fastaudio are the waty to g