tbh tacotron is named after the food

not that names matter much^^

ooh fastspeech has a cool idea > Specifically, we extract attention alignments from an encoder-decoder based teacher model for phoneme duration prediction, which is used by a length regulator to expand the source phoneme sequence to match the length of the target mel-spectrogram sequence for parallel mel-spectrogram generation.

So apparently trimming the white space is best with forced alignment according to sir Cookie

Because if you can map out when the words are spoken to the right phonemes you can just cut out all the white space before that first phoneme is spoken

Instead of guessing the volume and so forth

ohh yeah that makes sense

plus i'd imagine it'd help with some mis-transcripts (e.g. from telltale games) where the voice actor says some extra things before or after the text

Yeah so far the implementation I’ve seen is Montreal Forced Alignment

I’ll have to look into it but you create a dictionary with all phonemes

And see if each word spoken is in that dictionary

@hecko why do you have a custom data loader?

because i wanted to use /mlp/ datasets and the one collection thereof that i could find had a custom format

though i'd say its most useful aspect is that one can set it up and share the notebook with an 8th-grader and go "just do

run all

and google drive thing and it'll train"

Is it a data loader for breaking up the wav and txt files into certain directories?

yeah pretty much

Is there a better way to not keep gettign the K80 everytime? I've reset the runtime 50 times and keep getting k80

get Colab pro

pay for colab pro

😤

king koopa sounds like he's very, very sick

Yeah that needs some work

better

My model got that too.

my various starlight models, in order: [cw profanity] - 0 - forgot to filter noisy data - 1 - mild attempt to fix 0 by giving it just clean data - 2 - forgot to clip start silence (same applies to 0 and 1) - 3 - the "good" one

also finally tested out my old experiment of whether i could get the cruise seals from a hat in time to transfer-learn

r

and

l

(they cannot)

Wight wed. Light red. Woadwowwew. Roadroller.

also also the elec man model i trained without knowing about justinjohn training it, at 50 100 150 epochs;

audio

is the uberduck one (sounds muffled because it doesn't upscale the audio unlike the notebook) [again cw profanity at the end]

Nice results!

ehh