generic_v0.4

Trained for 98k steps on german common voice dataset.

vocoder_v0.1

WaveRNN vocoder trained for 142.000 steps. Can be used instead of griffin-lim algorithm, might deliver better results but requires more ressources to apply.

jonas_v0.1

Pre-trained German synthesizer model based on tacotron.

generic_v0.3

Trained for 75k steps on high quality voice.