A stacked convolutional neural network framework with multi-scale attention mechanism for text-independent voiceprint recognition

Abstract

Short-utterance speaker identification is a difficult area of study in natural language processing (NLP). Most cutting-edge experimental approaches for speech processing make use of convolutional neural networks (CNNs) and deep neural networks and analyse data in a unidirectional stream of time. In the past, approaches for identifying speakers that utilised CNNs often made use of highly dense or vast layers, leading to a large number of factors and significant computational expenses. In this article, we provide a novel multi-scale attention-focused 1-dimensional convolutional neural network (MSA-CNN) for recognising speakers that combines L1 and L2 norms. The multi-scale convolutional training architecture was developed to autonomously extract multi-scale characteristics of raw audio data by employing a variety of filter banks. In order for the multi-scale system to emphasis on important speaker feature characteristics in varying settings, a novel attention mechanism was built. In the end, it was combined and applied to the suggested multi-layered convolutional neural network framework to identify the speakers' labels. The recommended network model was tested on a number of standard voice databases and real time recorded corpus. The findings from the experiments demonstrate that our methodology outperformed a baseline CNN scheme (without an attention mechanism) in addition to conventional speaker identification techniques involving feature engineering, achieving an accuracy rate of 97.94% across numerous databases as well as distortion constraints.