too stupid

All good it has to do with the MLP dataset but I’m just confused as to some of the processes he’s doing

theres 250 epochs - am i just missing some phonemes or whats the situation?

those gaps must be when the speaker pauses for the next sentence

oh, brilliant - i'll get to work on some synthesis

thanks for the heads up

Here's how the training results look now -

Wonder if I should mess with the vocoder a bit

The synthesizer's pretty decent, but the vocoder seems to scuff things a bit

You can get super distinct results based on what you seed the encoder with

It learned 'pissed off narrator'

And here's seeding it w/ an emotional line

Added dependencies to Talknet Training and its now working for me, make sure your audio is wav and at 22k. I added code that theoretically should convert audio to wav at 22k if it isnt already but I haven't tested it. https://github.com/corranmac/TTSCollabs/blob/main/Working_TalkNet_Training.ipynb

python
      subprocess.run("ffmpeg -i %a -acodec pcm_u8 -ar 22050 %b" %a,b)

are you sure that's the right format? pcm_u8 is unsigned 8 bit isn't it? and these scripts all use 16 bit signed

might need lower learning rate to align across the gaps or more data

After vocoder training - Is it improved? I might've overtrained the vocoder..

trained the cruise seals from a hat in time because a friend requested it did only 150 epochs because there was like 4 minutes of audio data, so yeah it turned out not very well

currently training another one to experiment, the differences being: 1) 250 epochs b) all `r`s and `l`s turned into `w`s so there's a chance it carries over the actual `r`s and `l`s from the pretraining

oof

can someone check the model

1gRhOanrMJBVEsEIuV8HXOQHI1yXHEOO1

I'm not sure if this is the right place to ask, but when transcribing a speaker with awkward pauses is it best to write ellipses, a comma, or nothing? (feel free to ping me if you reply)

I have no idea I'm afraid

nothing is definitely bad, unless you want random awkward pauses in your output ellipses i've heard can be confused for periods so i use commas

okay, commas it is! thank you so much hecko!

I hope your harvest is bountiful and your days many!

well THAT seems like a scam bot

definitely overfit

So after I do a batch of 14, how do I do the next batch of 14(15-28) or am I misunderstanding something?

the batch size is just how many files it trains on at once in each iteration what you're looking for is the epoch count, where each epoch is 5 iterations i believe i don't see a built-in way to train an existing model for more epochs but there might be one