Problem with generated audio from pre-trained checkpoints

[swamiviv]

I used the pretrained checkpoints (64md_8k) for the sc09 dataset and generated samples as recommended. I used the following to read it and listen:

fname = 'commands_listen.mat'
mat = scipy.io.loadmat(fname)
import IPython.display as ipd
sr = 22050 # sample rate
ipd.Audio(mat['reconstructed'][0, :], rate=sr) # play a NumPy array

1. I find that most samples are illegible, but I can find some sounds here and there. Is that normal?
2. Out of curiosity, are the examples you present in the website cherry-picked?
3. Am I doing something wrong in generating the samples?

[andimarafioti]

1. I don't think that is normal
2. Not cherry picked, it's an array of sounds, you might notice some sound better than others.
3. Most likely you are doing something wrong, can you give more details?

[andimarafioti]

Are you using the speech checkpoint? there are a few checkpoints for piano too

[andimarafioti]

Could it be that you are indexing the matrix wrong? So mat['reconstructed'][:, 0] instead of mat['reconstructed'][0, :]