GM-LPC Based Multiband Analysis and Enhancement of Pathological Voice

An automated method for enhancing and repairing pathological voice can be considered a potential clinical tool in healthcare. Aiming to improve the intelligibility of pathological voices, this paper designs and implements a multi-band analysis and enhancement algorithm based on Gammatone Mean cosine similarity-Linear Predictive Coding (GM-LPC). First, the Gammatone filter bank is employed to simulate the auditory characteristics of human ears for multi-band analysis, achieving hierarchical frequency division processing of voice data. Then, MCS parameters are introduced to quantitatively measure the differences between normal and pathological voice spectra and track the changes in the pathological voice spectrum. Finally, The LPC analysis method is used to cross-synthesize voice signals with normal voice as a reference, completing the enhancement process and better preserving the individual voice characteristics of patients. The experiment demonstrates that the algorithm presented in this paper achieves objective evaluation scores of 0.570 and 2.913 for STOI and PESQ, respectively. Subjective analysis from the perspectives of time-domain and frequency-domain shows that the enhanced voice signal has less distortion, exhibits more concentrated and clear spectral characteristics, and significantly improves the quality of the voice signal.