Any tips on fixing alignment problems

[utunga]

Hello ! Training a model using tacotron. Pretty much just first crack at this (with coqui) but wondering if there are tips on avoiding alignment problems.

Voice sounds vaguely human after a few hundred epochs, but alignment is looking like this and I think this is reflected in the sound...

Just wondering if anyone made a checklist or something of common things to check with regards to bad alignment?

--

My audio data has been manually broken up (by humans) into utterances. (I have about 750 or so utterances total at the moment, but expecting more.)

One thing that i could do - though it would take a bit of work - is to run the whole data set through another model i already have that would allow me to clip the utterances more precisely around the spoken words. Would that be likely to help? (It's a non trivial extra step).

Alternatively I am wondering if I should maybe turn off the 'do_trim_silence' - these recordings are pretty good quality and made in a studio so perhaps 'trim silence' is not necessary? Current settings as follows

[utunga]

Any help much appreciated in advance.

[erogol]

https://tts.readthedocs.io/en/latest/faq.html

[utunga]

fwiw, the problem was that when I cloned the repo to add our dataset I accidentally /stupidly forked from the 0.0.15 tag (or something) instead of 0.1.x .. (bigger number bias) .. once I had the latest version of everything it all made a lot more sense

[kin0303]

Hi, have you got the answer?