Dont put in stereo

Generating Mels 84% 67/80 [00:01<00:00, 31.02it/s] /content/tacotron2/utils.py:14: WavFileWarning: Chunk (non-data) not understood, skipping it. sampling_rate, data = read(full_path) --------------------------------------------------------------------------- RuntimeError Traceback (most recent call last) in () 1 if generate_mels: ----> 2 create_mels() 3 frames /content/tacotron2/stft.py in transform(self, input_data) 82 83 # similar to librosa, reflect-pad the input ---> 84 input_data = input_data.view(num_batches, 1, num_samples) 85 input_data = F.pad( 86 input_data.unsqueeze(1), RuntimeError: shape '[1, 1, 51652]' is invalid for input of size 103304

I'm still getting the error

Again dont put in stereo

Convert to mono

I thought everything is mono

Put the dataset up here ill tale a look

So i can't find pop-80.tar

something's fishy

zip not tar

No, I mean the one on my drive

General question

For the transcript can the training notebook pick up **...**’s or is it only just periods, commas, exclamation points, and question marks?

there is a limited char set tacotron uses

but ' are included AFAIK

Alright I see, thank you!

this plus the space character are all supported

Alright cool thats good to know, didn’t know parentheses and such were also supported

I added this list to my tacotron FAQ

@User?

decrease the batch size

you might need to kill the colab runtime and start over to clear the GPU memory

even a batch size of five fails

you might need to kill the colab runtime and start over to clear the GPU memory

terminate?

yes

oh now it's working

thanks