Usage of audio_slice_frames, sample_frames, pad about universalvocoding

[bshall]

Hi @macarbonneau,

No problem, I'm glad you found the repo useful. I haven't tried using the end (or beginning) segments but there's no real reason it shouldn't work. The thinking behind using the middle segment was to match the training and inference conditions as much as possible. At inference time most of the input to the autoregressive part of the model (rnn2) will have context from the future and the past. So taking the middle segment is "closer" to what the network will see during inference. If you used the end segment, for example, the autoregressive component wouldn't have future context at training time and the mismatch might cause problems during generation.

Hope that explains my thinking. If anything is unclear let me know.

One of the negative side effects of only using the middle segment is that there are sometimes small artifacts at the beginning or end of the generated audio. For the best quality it might be worth putting in some extra time to train on the entire segment.

[bshall]

Hi @wade3han,

Yeah, I should add some comments explaining those parameters.

First, sample_frames is the number of frames sampled from the melspectorgram that get fed into the conditioning network (rnn1 in the model). The output then gets upsampled and sent to the auto-regressive part (rnn2 in the model). But if we set sample_frames to 40 then after upsampling there are 40 x 200 = 8000 samples which takes far too long to train.

To speed things up I only take the middle audio_slice_frames, upsample them and then use that to condition rnn2. The pad parameter is just how many frames are on either side of the middle audio_slice_frames. So for the default config this would be (40 - 8) / 2 = 16 frames. To account for only taking the middle frames I also padded the melspectograms by pad on both sided in preprocess.py.

I hope that helps.

[wade3han]

Thanks for your reply!

I guessed that strange artifacts like below happens because of those hyperparameters. Don't you have seen those artifacts? I got those artifacts mainly on the front and back of audio files.

[bshall]

No problem.

Are you using the pretrained model and the generate.py script? Also, what input audio are you using? Is it something from the ZeroSpeech corpus or your own?

I'm not getting any of those artifacts. For example, here's the original audio:

And here's the reconstruction:

[wade3han]

Well, I was training new model from scratch using Korean speech data corpus. It has 300 hours amount of various speakers' utterances, and I was getting those artifacts after I tried to use audio_slice_frames=16 instead of 8. I believed using bigger audio_slice_frames can help training.

Actually, I'm not sure why those artifacts are generated... I will share you if i figure out why. Please share your opinion if you have any ideas.

[dipjyoti92]

@bshall I have one question to your first reply in this thread. Instead of having 40 mel frames, why not use 8 mel frames itself in the input of the rnn1 layer itself?

[bshall]

Hi @dipjyoti92, sorry about the delay I've been away. I found that using only 8 frames as input to the rnn1 layer results in the generated audio being only silences. I think 8 frames is too short for the rnn to learn to appropriately use the reset gate although I haven't investigated this thoroughly.

[macarbonneau]

Hello @bshall !
Thank you for the awesome repo. Your code is very clean, I'm impressed. I'm playing a bit with your implementation and I have a question. Why do you take middle of the mel segment? Why not just the end? is there a benefit of having the padding at the end?

Thank you!!