<b>An Intelligent Speech Recognition Framework Using Hidden Markov Models and Actor–Critic Deep Reinforcement Learning for Low-Resource African Languages</b>

Abstract

ASR also plays a crucial role in assistive technologies for individuals with disabilities by enabling them to manage their surroundings more effectively through dialing phone numbers, operating light switches, and controlling home appliances, thereby contributing to the development of smart home systems. This study extracts features from isolated speech using Mel-Frequency Cepstral Coefficients (MFCC) and Bidirectional Long Short-Term Memory (BiLSTM) networks to ensure speaker invariance and enhance feature localization. Deep learning techniques were employed to explicitly normalize speech spectral features. Numerous pattern recognition and regression tasks have demonstrated the effectiveness of LSTM-based architectures. The novelty of this study lies in the integrating a hybrid MFCC–DNN–HMM framework to achieve high speech recognition accuracy for isolated words. The model achieved an accuracy of 0.945 (94.5%), indicating that it correctly classified the majority of instances. The precision obtained was 0.901, meaning that 90.1% of the instances identified as positive were correctly classified. The recall rate was 0.92, indicating that 92% of the actual positive instances were successfully detected by the system. The F1-score was 0.909, reflecting a balanced measure of precision and recall.