Automated-Speech-Recognition-System-for-Spoken-digits-Using-Deep-Learning

Deep Learning, Convolutional Neural Network,Regularisation, Convolutional Neural Network with Residual Connection Designed, implemented and evaluated an Automated Speech Recognition (ASR) system for spoken digits using Deep Learning methods. Extracted audio-based spectrograms from the raw .wav files of speech using digital signal processing in python. Convolutional Neural Network(CNN) modes have been explored. Three models have been applied $-$ first with $4$ CNN blocks with 512 filters with 69.68 percent accuracy, Second with $4$ CNN blocks with 512 filters regularised with 72.42 percent accuracy and the last one is the CNN model with Residual Connection with an outperformance of 74.64 percent. Demonstrated training/validation performance with visualisation during the hyperparameter optimisation in training. Demonstrated performance using the confusion plot for each task.