Deep features-based dialect and mood recognition using Assamese telephonic speech

Abstract

Learning aided methods are popular for designing automatic speech recognition (ASR) systems. Majority of works have used shallow models in combination with mel frequency cepstral coefficients (MFCC) and other features for speech recognition applications. Although these shallow models are effective but incorporating deep features in the mechanism for speech processing applications is necessary to increase the efficiency. Despite of considerable amount of works on the design of deep learning topologies and training paradigms in supervised domain, very few works have concentrated on deep features which are essential to capture detailed information of speech. This work focuses on the generation of deep features using stacked auto-encoder for normal and time shifted telephonic speech samples in Assamese language with mood and dialect variations. Experimental results show that the deep features learned by the stacked auto-encoder performs better while it is configured for Assamese speech recognition with mood and dialect variations.