Feature based Speaker Embedding on conversational speeches

Abstract

Identifying the speaker of a specific speech by examining the speech features of the voice is called speaker identification. The task of speaker identification consists of three main phases which are feature extraction, feature embedding and voice classification. Speaker embedding is the process of modeling the voice of a person where the model of the utterance can uniquely represent the speaker of that voice. Speaker embedding is a commonly used method in Automatic Speaker Recognition systems to identify the voice of the speaker. Currently, Deep Neural Networks based approaches are used in these systems for speech feature extraction and speech embedding. The performances of different approaches heavily depend on the noise factor and suitability of selected features of the audio data. MFCC, LPC, Dimensional filter banks are some of the frequently used speech features in speaker recognition. This speaker recognition research focuses on the usage of speech features for speaker embedding that are fitting for the speaker identification in conversational environment using a Convolutional Neural Network based approach.