Using Variational Autoencoders to Reduce Noise in Speech Processing for Cochlear Implants

Abstract

In this thesis, I present a variational autoencoder model to reduce noise in speech processing for cochlear implants. The model was trained on noisy and clean speech samples that were converted to spectrograms, to be able to create clean speech from a noisy speech input. The variational autoencoder consists of a convolutional encoder and a deconvolutional decoder. Furthermore, a cochlear implant simulator was developed to allow for testing on normal hearing participants. Participants were asked to judge the quality of original noisy speech data and noisy speech data that was processed by the model, where both of which went through the cochlear implant simulator. Results indicate that the quality of speech was worse when it had went through the variational autoencoder compared to the original noisy speech.