Speech Recognition Using Multi-Deep Learning Techniques for Person Identification

Abstract

Speech recognition is one of the fastest-growing technologies. It offers multiple potential benefits and has varied applications in numerous sectors. Many individuals are incapable of conversing with one another due to language barriers. We intend to reduce this barrier through our project, which was designed and developed to achieve solutions in specific settings that considerably assist individuals in sharing information by using voice input to control a computer. It is possible to do speech recognition tasks after gaining knowledge in a variety of disciplines, including linguistics and computer science. It is not a solitary endeavour. Speech recognition techniques include HMM (hidden Markov model), DTW (dynamic temporal warping)-based voice recognition, neural networks, deep feedforward and recurrent neural networks, and end-to-end automatic speech recognition. Different neural networks utilized for automatic voice recognition are the major topic of this study. In this work, a convolutional neural network (CNN) is used in an effort to develop a system that can identify a human speaker (CNN). This study employs the mel-frequency cepstral coefficients (MFCC)-CNN and raw waveform (RW)-CNN techniques. The first approach is a traditional one that makes use of MFCC and the audio features, which are then entered into CNN to carry out a process. The suggested CNN's training process will start by accepting input in the form of a picture. The second approach, RW-CNN, follows the identical processes as the first approach but bypasses the MFCC phases in favour of straight admission to CNN. Both techniques employed the same CNN structure. In this study, both RW-CNN and MFCC-CNN achieved an accuracy of 96%.