it did on the start, but its still a bit off, and i thought it will go on untill the "gun smoke - gun smoke" part, it became mumble but i fixed it making its "Throw you in a ChokEGuNSmokGGunsmok." instead "Throw you in a choke. gun smoke gun smoke"

aight last, this is the "before using synthesizer"

and "after using synthesizer"

Do you guys trim your audio in any way. Like remove silence from front and end of audio?

yes. i have my processed source (normalized and vocalremover5'ed if needed) on audacity and use a label track to trim the silence out

if its already clipped and it has silence try using sox

clipped?

basically already separated to different wav files

ok i'll try using sox thanks for that suggestion

How does it determine where they start speaking?

Just the increase in volume?

yes

I only trim my clips if they get too long, then I split 'em up and use periods at the end of the first one that's split

this article gives a pretty good explanation of the sox silence filter

Thanks!

I've also seen in Cookie's repo that he does (or used to do) a trim pass has anyone applied something similar?

I'm not sure if this is necessary or the reason for it

i dont think its needed because ive never come across that before

if i remember correctly trimming silence lets the model learn attention (see: tacotron2 issues tab)

It starts from this line

He applies high pass, low pass filter, resamples to target sample rate, and trims

I'll skip the filter part and trim pass for now thanks @User

Does anyone here use Montreal Forced Aligner to get the phonemes?

my notebook has a silence trimmer, specifically for the /mlp/ datasets because those have silence around them

Yeah I’m planning to use sox to trim the edges

hadn't thunk of frequency filtering though

speaking of /mlp/ i wonder how worth it is to train on a half-clean half-noisy dataset first, then train some more on just the clean one like would the extra data outweigh the neurons wasted on replicating the noise