This repository contains the code to reproduce the core results from the paper "Learning Latent Representations for Speech Generation and Transformation".
Great thanks for sharing the code. I have a question about the range of log-likelihood to expect with respect to TIMIT data. I am getting inbetween range(-200,-400) for Filter bank features. Thank you for your help.
Hi, I read your paper and it said the sequence of features of each sentence is splitted into 80 (or 200) frame chunks. Each chunk is an input "image" for the Convolution VAE. I guess for each sentence, the number of chunks is equal to the number of frames. I'm looking for the script where you split each sentence into "images" but I couldn't find it in the repo. I know it can be a simple function but it will be very helpful for me. Where can I find the part in your repo?
Thanks