It sounded off

Why did it sound like that

This was the transcript for my first model

Any thoughts?

theres like a million and one points of failure a transcript isnt gonna make many obvious

I'm so confused what???

hiii sorry to barge in but piggybacking off of dawn it could be because of the quality of wav files, possibly you didnt see the sub guide or something and many more

and i mean many more

Effects could very much affect it. (Ex: Current upload of Imogen Heap, if it's still up)

and also background noise

the model couldve just been undertrained

the params couldve been bad

the transcript couldve had errors

the cadence couldve fluctuated a bit too much

honestly yeah it could be undertrained

How was I overtraining a model?

Nobody's saying you did but it's entirely possible

They said undertrain

Like how should I train a model in order for it to be “undertrained”?

stop training before a good val loss

What’s a val loss?

Validation loss. > Epoch: 1 Validation loss 388: 0.550257

How close a model is to the ground truth. Probably not correct but eh

Also maybe I shoulda just used all the files to begin with if thisis already epoch 1

I feel like this is the part where its undertrained

Each time I open a file on Audacity it’s raw and all static

try installing this and then relaunch Audacity

I got 200 epochs for my Itachi model using Tacotron 2 legacy. Is it supposed to take this long or should I do something?

AssertionError Traceback (most recent call last) in () 73 ) 74 x_f0s = f0s[x_name].view(1, -1).to("cuda:0") ---> 75 x_spect = model.force_spectrogram(tokens=x_tokens, durs=x_durs, f0=x_f0s) 76 rel_path = os.path.splitext(x.split("|")[0].strip())[0][5:] 77 abs_dir = os.path.join(outdir, os.path.dirname(rel_path)) /usr/local/lib/python3.7/dist-packages/nemo/collections/tts/models/talknet.py in force_spectrogram(self, tokens, durs, f0, **kwargs) 360 text_len = torch.tensor(tokens.shape[-1], dtype=torch.long).unsqueeze(0) 361 durs_len = torch.tensor(durs.shape[-1], dtype=torch.long).unsqueeze(0) --> 362 assert text_len == durs_len 363 364 # Spect any one know what that this is