wavenet_vocoder_liepa

Contents for

View the Project on GitHub aleksas/wavenet_vocoder_liepa

Train new model from start conditioning on 9 speakers. As a result one can understand the spoken language but disturbances are too obvious. Additionally speaker 12: D17 had originally too high noise level that was not cleaned up automaticcaly, thus has to be replaced with another speaker.

Experiment evaluation results

Following utterances are generated using evaluate.py script. First raw represents predicted utternace and the second is ground truth target.

Speaker 7 utterance 0

Speaker 12 utterance 2

Speaker 4 utterance 3

Speaker 11 utterance 4

Speaker 3 utterance 5

Speaker 15 utterance 10

Speaker 17 utterance 17

Speaker 1 utterance 23

Speakers

The speaker indeces correspond to speker directories in Liepa database in order specified in dataset (See var. available_speakers). Notice that only speakers with sentence utterances were considered.