# ESPERNet
ESPERNet is a set of AI models for audio-to-audio speech processing. The versions available here have been trained on the Mozilla CommonVoice dataset.
**This model is still in development! Weights will be uploaded as soon as hyperparameter tuning and training are complete.**
ESPERNet is built from three models: an encoder, a decoder, and a classifier.
Together, they form a VLGAN architecture.
***(Lee, Je-Yeol & Choi, Sang-Il. (2020). Improvement of Learning Stability of Generative Adversarial Network Using Variational Learning. Applied Sciences. 10. 4528. 10.3390/app10134528. )***
The VAE latent space is split into two parts: a low-dimensional, time-dependent phoneme space and a larger, time-constant style-and-speaker space.
This difference in dimensionality enables the encoder to separate speech content from speaker identity.
***(Xie, Yuying, et al. "Speaker and style disentanglement of speech based on contrastive predictive coding supported factorized variational autoen…
