Person identification based on voice recognition

Abstract

TIn this era of technological advancements, speech recognition has gained prominence and increased demand especially with the rise of AI based assistants like Google Assistant, Microsoft Cortana, Apple's Siri, Amazon's Alexa to name a few. Identifying an individual using spoken words and phrases automatically is referred to as speaker recognition. In this paper, a Convolutional Neural Network (CNN) is used for the task of person identification from the audio database. The dataset being used is a self-made database of 60 subjects by recording voices. Two networks namely, Inception v3 and MobileNet v1 are used for the task of speaker recognition by using spectrograms of audio database. A dataset of 12000 spectrogram images is used for this task of which, 9600 images are used for training, 1200 images for validation of the networks, and 1200 images are used for testing. Inception v3 network outperforms the MobileNet v1 network by providing an accuracy of 85.5%.