search

fatchord / FFTNet

Pytorch Implementation of FFTNet
86 stars 20 forks source link

Specific audio generation #4

Open gcunhase opened 5 years ago

gcunhase commented 5 years ago

Hi, is it possible to use this model conditioned on the first few samples to generate a specific audio? Say I want to generate audio1.wav, then after training with my dataset, I'd be able to produce that audio given the first N samples.

Thank you for your time

fatchord commented 5 years ago

@gcunhase You mean like how you would seed an RNN? Anyway, I would be skeptical - mainly because this is a convolutional model with no hidden state - it doesn't 'remember' anything outside of its receptive field.

gcunhase commented 5 years ago

@fatchord Yeah, I mean it like an RNN, like seeding the model with an initiator so it produces a sample that we want.

fatchord commented 5 years ago

@gcunhase I don't think it would work without conditioning the model with side information. Once the initial seed moves out of the receptive field it'll just generate random stuff that it's seen during training.