A Hybrid Architecture Based on Conformer and CRN Parallelism for Regulated Speech Quality Measurement Model

Assessing the quality of speeches between pilots and air traffic controllers(ATC) is a crucial challenge for intelligent and industrialized civil aviation. Assessing the voice quality between air traffic controllers and pilots during communication requires both clean speech without noise and speech with noise introduced. However, clean speech without noise cannot be obtained. To tackle this issue, we propose a non-intrusive pilots and ATC speech quality evaluation algorithm that operates in parallel with a conformer network and convolutional recurrent network (CRN). To more realistically characterize the key factors impacting speech quality in pilots and ATC calls, our algorithm initially employs Meer cepstrum coefficients (MFCC) as the characterization method for speech signals. Our algorithm constructs a convolutional neural network to extract features from the MFCC, and then utilizes a conformer network and a CRN network to simultaneously extract the relevant features of speech quality from the pilots and ATC speech, using global attention adaptation and inter-frame time dependence methods, respectively. Finally, the features are mapped to objective scores using attention pooling methods. Experimental results based on an open-source database indicate that our proposed algorithm correlates exceptionally well with the mean opinion score (MOS). By utilizing this algorithm to measure ATC speech, we can effectively distinguish pilots and ATC speech data with different speech quality. Overall, our nonintrusive ATC speech quality evaluation algorithm provides a promising solution to the challenges facing intelligent and industrialized civil aviation.