This is the PyTorch implementation of the SDM-VAE model proposed in [1]. The codes are inspired by this repository.
The standard VAE and SDM-VAE models are provided in ./model
.
Set the training properties (e.g., network architecture, STFT parameters, etc.) in the config
files provided inside ./configuration
. The training then proceeds as follows:
# Train a VAE model:
python train_model.py --cfg ./configuration/cfg_VAE.ini
# Train an SDM-VAE model:
python train_model.py --cfg ./configuration/cfg_SDM_VAE.ini
You can evaluate a trained VAE model in a speech analysis-resynthesis task via test_speech_analysis_resynthesis.py
.
If you have any questions, please feel free to reach me out via email: mostafa[dot]sadeghi[at]inria[dot]fr
[1] Mostafa Sadeghi and Paul Magron, "A Sparsity-promoting Dictionary Model for Variational Autoencoders," in Procedeeings of Interspeech, September 2022.