Music/Speech/SFx discriminator model trained on GTZAN and BBC SFx

Fixed some bugs on the training code and trained on the CPU using a number of 4 epochs. On the same dataset it gives much higher f1 scores than before:

Type	precision	recall	f1-score	support
Music	1.00	0.97	0.99	713
Speech	0.98	0.98	0.98	527
SFx	0.96	0.99	0.98	494

Music/Speech/SFx discriminator model trained on GTZAN and BBC SFx

.h5 and .json files.

Music/Speech/SFx discriminator model trained on GTZAN

Adapted VGGish trained on GTZAN music/speech discriminator dataset and augmented by samples from the BBC SFx library