Multidimensional feature diversity based speech signal acquisition

Abstract

Computer intelligence methods are the buzz of identifying the state and context of diversified domains. The computer intelligence methods mostly fall into the domain of machine learning, which often trained by the known data of the domain context and state. These methods even spanned significantly to the diversified contexts of speech domains such as speech recognition, emotion recognition, speaker recognition, and many more. However, the contemporary methods of speech acquisition or speech separation have majorly relied on filters and other signal processing techniques, which are considerably underrated to identify the speech signal that exists among the highly correlated noise signals. In this context, the contribution of this manuscript is a machine learning mechanism that intended to identify the speech signal bounded with highly correlated noise signals. The method proposed is using the diversified features extracted from the signals representing the speech and other correlated noises. The experimental study has carried on the signal and correlated noise utterances collected from the benchmark dataset CHiME-5, which has meant for machine learning-based signal processing. The performance analysis of the proposed model has scaled by comparing it with the contemporary model.