Because each of them uses the speech parameter generation algorithm with a different setting. HMGenS uses a better, but computationally expensive one.
Thank you.But it made me confusing. Because I synthesized a sentence with two ways: 1. HMGenS_SPTK and 2. hts_engine.For the former, every time I synthesize, I perform HHEd to make unseen models contain in the sentence, then use the model set to do synthesis.For the latter, I don't perform HHEd to make unseen models contain in the sentence. Just convert the original HTS engine model set mmf file to HTS engine format, and then do synthesis.The generated voice from these two ways were different. The generated voice of the latter one contained some explosive sound noise. I thought it was because HTS engine didn't handle unseen model. Now I'm confusing why these two generated voice were different.2012/5/2 Heiga ZEN (Byung Ha CHUN) <heigazen@xxxxxxxxxx>
2012/05/01 18:52 "Kwan Lisa" <lisakwan1102@xxxxxxxxx>:
> Does how HTS engine handles unseen models the same way as HHEd does?
> 2012/5/2 Heiga ZEN (Byung Ha CHUN) <heigazen@xxxxxxxxxx>
>> 2012/5/1 Kwan Lisa <lisakwan1102@xxxxxxxxx>:
>> > I want to know how to implement a real time TTS system that can deal with
>> > unseen models. I trained a set of models and converted to HTS-engine format
>> > so that it can easier to synthesize the arbitrary input sentence. But I
>> > don't know how this could deal with unseen models. Unseen models always
>> > exist for a given set of models. If you want to perform HHEd to make the
>> > unseen model to deal with it and then convert it to HTS-engine format, it
>> > could the very slow for a real time TTS system because HHEd takes a lot of
>> > time. Is there any effective way to deal with unseen models?
>> hts_engine does handles unseen models. It traverses decision trees
>> and finds leaf nodes given labels. Then it synthesizes speech.
>> Please check the tree-related source codes in the hts_engine release.
>> Heiga ZEN (in Japanese)
>> Byung Ha CHUN (in Korean)