Speaker Differentiation Using a Convolutional Autoencoder

摘要

In this work, a deep learning solution for differentiating speaker voices in audio given two microphone sources is presented as a step towards solving the cocktail party problem. A convolutional autoencoder was trained using a small sample size of data to associate audio snippets with categorical labels. Audio snippets collected as part of this work were used for training and evaluating the model. Audio was converted to mel-frequency cepstrum representation prior to classification. The collective processed data was labeled according to the person or collection of persons speaking. The model was trained and evaluated using data with two, three, four, five, and six categories. The result was a model that recognizes when different people are speaking in a 2-person, 3-person, 4-person, 5-person, and 6-person conversation with an accuracy of 99.29%, 97.62%, 96.43%, 93.43%, and 88.1%, respectively. Experimental comparisons between the five versions of the model are presented.