Awesome
Factorized Hierarchical Variational Autoencoders
This repository contains the code to reproduce the core results from the paper Unsupervised Learning of Disentangled and Interpretable Representations from Sequential Data.
To cite this work, please use
@inproceedings{hsu2017learning,
title={Unsupervised Learning of Disentangled and Interpretable Representations from Sequential Data},
author={Hsu, Wei-Ning and Zhang, Yu and Glass, James},
booktitle={Advances in Neural Information Processing Systems},
year={2017},
}
Dependencies
This project uses Python 2.7.6. Before running the code, you have to install
The former 6 dependencies can be installed using pip by running
pip install -r requirements.txt
The last one requires Kaldi before a specific commit (d1e1e3b). If you don't have
Kaldi before that version, you can install both Kaldi and Kaldi-Python by running
make all
Usage
The code structure follows Kaldi's convention. Scripts for each dataset can be found in
egs/<dataset> folder. If you have questions, please write an email to wnhsu@csail.mit.edu
TIMIT
To reproduce the experiments for TIMIT, run:
# stage = -1 for data preparation; run only once.
cd egs/timit
./run_fhvae.sh --TIMIT_RAW_DATA <timit_raw_data_dir> --stage -1
Aurora-4
Before running FHVAE experiments, you need to have the following Kaldi data directories
of Aurora-4 prepared:
<aurora4_kaldi_egs_dir>/data/{train_si84_clean, train_si84_multi, dev_0330, test_eval92}.
To reproduce the experiments for Aurora-4, run:
# stage = -1 for data preparation; run only once.
cd egs/aurora4
./run_fhvae.sh --AURORA4_KALDI_EGS <aurora4_kaldi_egs_dir> --stage -1
Experimental Results
A. Visualizing Factorization
We show the combination of 6 latent sequence variables and 6 latent segment variables. The latent sequence variable for each column is the same; likewise, the latent segment variable for each row is the same.
TIMIT (alpha = 0)
<img src="img/fac_timit_alpha0.png" width="500"/>TIMIT (alpha = 10)
<img src="img/fac_timit_alpha10.png" width="500"/>Aurora-4 (alpha = 0)
<img src="img/fac_aurora4_alpha0.png" width="500"/>Aurora-4 (alpha = 10)
<img src="img/fac_aurora4_alpha10.png" width="500"/>B. Utterance-Level S-Vector Modification
We show replacing S-vector of a target utterance with that of a reference utterance to alter the sequence-level attribute, which can be applied to speaker conversion or denoising. The FHVAE model used for following modifications is trained on Aurora-4 with alpha = 0