Isolated Word Speech Recognition Using Convolutional Neural Network

This research aims to design and develop an accurate speech recognition system for a set of predefined words collected from short audio clips. It uses The Speech Commands Dataset v0.01 provided by Google's TensorFlow. Isolated word speech recognition can be implemented in voice user interfaces for applications with key-word spotting. The end goal is to classify and recognize ten words, along with classes for “unknown” words besides the “silence” class. The problems that face the current speech recognition technology like the acoustical noise and variations in recording environments are also solved and addressed here. To extract useful information from the signal, two methods of feature extraction were used: MFCCs and Mel-spectrograms. For classification, the convolutional neural network (CNN) was used. Different models were developed for this research, where each model has different architecture (1D-convnet and 2D-convnet). During training, techniques like batch normalization, regularization, and dropout were added to improve the accuracy and maintain the efficiency of the models. As a result of our experiments, The final model (2D-convnet with MFCC-16000) achieved an accuracy of 97.07% for training and 96.19% for testing.