A Non-Intrusive Speech Quality Assessment Method for Low-Rate Communication

Although PESQ (Perceptual Evaluation of Speech Quality) serves as an important benchmark for assessing narrow-band speech quality, its effectiveness depends on the availability of reference signals, which is not always feasible in practical applications. To surmount this limitation, numerous non-intrusive assessment methods have been proposed. However, most of them focus on wideband speech, with scant research dedicated to narrowband speech. To this end, this paper proposes a non-intrusive speech quality assessment model suitable for low-bitrate narrowband communication scenarios. The model integrates Convolutional Neural Networks (CNN), Bidirectional Long Short-Term Memory (BLSTM) networks, and a frequency band attention mechanism, effectively extracting key features from narrowband speech signals and conducting quality evaluation. Experimental results show that the proposed model significantly outperforms the baselines, achieving a Mean Squared Error (RMSE) of 0.0200, a Pearson Linear Correlation Coefficient (PLCC) of 0.955, and a Spearman Rank Correlation Coefficient (SRCC) of 0.954, demonstrating higher accuracy and robustness.