Automatic Speech Recognition: Comparisons Between Convolutional Neural Networks, Hidden Markov Model and Hybrid Architecture

Abstract

Automatic Speech Recognition (ASR) systems have been widely used as a practical method of interaction between humans and devices. They are typically employed to enhance the accessibility of devices and to improve the security of systems, among other purposes. However, the design of speech-based systems imposes many challenges due to their particularities. Currently, the majority of ASR systems is based on the Hidden Markov Model (HMM), and, more recently, on Convolutional Neural Networks (CNN). The present research evaluates the performance of Hidden Markov Model (HMM) and Convolutional Neural Network (CNN) algorithms in speech recognition and proposes a novel hybrid approach that combines both methods. The study assesses various performance metrics, including accuracy, precision, recall, F1-score, response time, and computational cost. The experimental tests show that the integration between HMM and CNN increased the accuracy by 6% and 8% when compared to HMM and CNN isolated, respectively, in accordance with results presented in previous papers. However, the results of the ANOVA test revealed that the difference in question is not statistically significant, and the HMM-only approach still being an interesting option for embedded systems due to its lesser demanded computational effort.