Speech Conversion Model Design for Non-Audible Murmurs NAM with Enhanced Accuracy Using Krill Herd Algorithm

Non-audible murmurs (NAM) pose a challenging problem for speech recognition because of their weak-intensity signals, resulting in high Word Error Rates (WER). In this study, an improved speech conversion model based on Generative Adversarial Networks (GANs) and Full-Rank Gaussian Mixture Models (FR-GMMs) is put forward, with additional optimization using the Krill Herd (KH) algorithm. The KH algorithm regularizes model parameters with improved speech clarity and low WER. Experimental outcomes indicate a 32.8% absolute reduction in WER for training data and a 50.1% relative improvement in test data using the GAN-based approach. In contrast, the FR-GMM model yielded a 4.7% reduction in WER. Outcomes indicate the efficacy of bio-inspired optimization algorithms to propel NAM speech recognition, providing access to superior silent communication programs and assistive speech technologies.