01/03/2018

The problem of voice conversion is an interesting one. It involves changing the speaker identity while keeping the content unchanged. I attempted a deep learning solution for this problem.  Using LSTMs Since audio data is sequential in nature and the knowledge of most recent values is essential in speech generation, I chose LSTM (long short term memory) networks for my implementation. I implemented an LSTM autoencoder which takes in features of the source speaker and optimizes the mean squared