| Speaker Identification (SID) is one of the most prominent and leading research arena that gained tremendous momentum in the recent years. This increased attention of machine learning approaches, especially deep neural networks (DNNs), further enhanced the interest towards SID research. Despite success stories of DNNs, there has been limited attention towards deep autoencoders (DAEs) and their applications has not been explored properly. In this paper, we attempt to address this gap by applying DAEs to identify speakers using analytical and experimental research prospects. The experiments were conducted using the data obtained from 84 speakers provided in AN4 corpus. To understand the significance of 'depth', i.e. the number of autoencoders in the DAE, multiple experiments with different number of autoencoder layers in the DAE were conducted. The experimental results show that DAE network with three autoencoders was able to achieve superior identification accuracy of 98.8% over the traditional neural networks. The findings of this study confirm the importance of 'depth' as highlighted in previous deep learning studies especially with the difference in accuracy between regular back propagation and layer-wise training. This paper further provides a new direction in the implementation of deep autoencoders for speaker identification. |
|
Date: Thursday, July 16, 2020 Language: English Downloded 9 times. |