Uberduck

my attempt at doing the overwatch hl2 voice but it's failing miserably

not a voice of a existing character, just recorded my own lines and edited them lmao

Well looks like patchy didn't work with 41 voice clips

I can't because I deleted the model because it was too bad

The same thing happened to me with the lu bu voice from dynasty warriors

happened to me time and time again with Vicky too

how does this even happen in the first place

I retrained patchy and he had some better results

Man my voices are just coming out worse and worse. Not sure if my expectations are just fucked, but it's like I get only one sentence to work before it strokes, or I get nothing audible at all.

doing the tacotron2 training again now with cmudict
already getting alignment image

i also thought of doubling the number of epochs since half of training goes to arpabet sequences and the other half goes to english words

`The {T AA1 K OW0 T R AA1 N} 2 and {W EY1 V G L OW1} model form a text-to-speech system that {EH0 N EY1 B AH0 L Z} the user to synthesize a natural sounding speech from raw transcripts without any additional prosody information.`

arpabet sequences provided by <http://www.speech.cs.cmu.edu/cgi-bin/cmudict>

model's currently broken rn as it strokes out after 1 or 2 sentences, but here's a preview of the next voice.

how did you get your tacotron model to use specific arpabet sounds

bc I know my own tacotron model cant differenciate between aa1 and aa2 and shit lol

i created a patch for tacotron2 that (randomly) replaces words in the dataset with cmudict equivalents as it trains

so it randomly converts words in the list to cmu artict equals with the arpabet sounds assigned to them?

when i packaged this i forgot to include merged.dict.txt 
this goes on /content/tacotron2/