[hts-users:04609] Re: Problem on Synthesis using DNN on HTS - demo - 2.3

[Ericson Sarmento <ericsonsarmento@xxxxxxxxx>]

Here are some synthesis samples using  HMM model by the same question file, and the same scripts.

2018-07-06 14:51 GMT-03:00 Ericson Sarmento <ericsonsarmento@xxxxxxxxx>:

Hello all,

While I was using HTS demo 2.3 which supports DNN training for duration modeling for my language (Brazilian Portuguese), I caught in a problem on generated synthesized sample audios which comes with an irregular duration of the (I guess) first utterance's phone (at all) and the other phones seem to be synthesized with the correct duration, but at the same time the synthesis from HMM model trained with the same scripts is OK at all. The question file is the same for train both models (HMM and DNN).

Someone ever gets this type of problem while using HTS demo 2.3?
What type of problem can be generating this distortion?
How can I do some troubleshooting about this problem?

I'm sending some samples of synthesis using DNN in this message for your appreciation.

Regards,

-- Ericson Sarmento

Attachment: gen_m001_0001.wav
Description: Wave audio

Attachment: gen_m001_0002.wav
Description: Wave audio

Attachment: gen_m001_0003.wav
Description: Wave audio

Attachment: gen_m001_0004.wav
Description: Wave audio