oh

I see

So each graph is for the next clip

and then rinse and repeat after all the clips

pretty much

Ok yeah I am still learning with all this... How much total audio time do you think I should have for a really good model, or is it more dependent on training time?

also the recommended length for clips is between 2 seconds and 11 seconds (with silence trimmed at the beginning and end)

a good model can depend on the amount of clips but consistency between the clips is another factor

if you get clips to around the hundreds you should be good (which corresponds to just a few minutes of data)

Yeah i currently have 75 clips from 3 youtube videos of william daniels, 2 interviews and one with him acting in an episode of boy meets world

Im trying to make a KITT voice which is actually really fortunate he didn't do any sort of accent for that just his normal talking voice same as everything else

if you are training smaller datasets batch size and/or learning rate could be reduced to avoid overfitting (which is when the model starts forgetting the base english youre actually training from in favor of your dataset)

Ok yeah that makes sense

I've got it down to 0.03 now

Gonna wait for it to either finish or time out and then try it out

thats a really good validation loss

how many epochs

Epochs done or the number i put in the config?

epochs done

Epoch 397

Batch size 10

astaghfirullah that is a lot for 75 wavs

the number i put in is 500

I am really good with computers but a complete newcomer to ml and ai stuff

usually i stop training when validation loss is below 0.2 (i let it go down to 0.1) and/or when i reach a certain amount of iterations

Yeah the tutorial said 0.15, someone else above the other day said it improves a lot at 0.08 so i thought what the hell i have time ill go as long as it wants to

If you want I can send you the wavs and transcript to get your opinion, like I said my first time doing this

Hey! I assume people have had this problem before but I couldn't find anything. I'm having this issue. I probably made a dumb mistake but what did I do wrong? here's the output.

Generating mels
Generating Mels
100%
100/100 [00:02<00:00, 38.26it/s]
/content/tacotron2/utils.py:14: WavFileWarning: Chunk (non-data) not understood, skipping it.
  sampling_rate, data = read(full_path)
Checking for missing files
Checking Training Files
 
[WARNING] does not exist.
---------------------------------------------------------------------------
IndexError                                Traceback (most recent call last)
<ipython-input-13-197f98831c23> in <module>()
      9 get_ipython().system("sed -i -- 's,.wav|,.npy|,g' {hparams.training_files}; sed -i -- 's,.wav|,.npy|,g' {hparams.validation_files}")
     10 
---> 11 check_dataset(hparams)

1 frames
<ipython-input-5-0682750d4349> in check_arr(filelist_arr)
    369             if (not os.path.exists(file[0])):
    370                 print("|".join(file), "\n[WARNING] does not exist.")
--> 371             if len(file[1]) < 3:
    372                 print("|".join(file), "\n[info] has no/very little text.")
    373             if not ((file[1].strip())[-1] in r"!?,.;:"):

IndexError: list index out of range

most likely an empty line in the transcript

ah alright