FRS: Adaptive Score for Improving Acoustic Source Classification From Noisy Signals

This letter introduces a Frame Relevance Score (FRS) to improve the classification of environmental acoustic sources from noisy speech signals. The importance of each short-time frame for the classification results is objectively interpreted by SHapley Additive exPlanations (SHAP) values. The FRS enables the selection of frames that are more appropriate to improve the discrimination power of the acoustic models. The FRS-based frame selection can be used as a pre-training strategy to any classification approach. Evaluation experiments consider the recognition of ten background sources from noisy speech signals. The classical system based on MFCC and GMM is adopted to prove that the selected frames can better distinguish the acoustic classes. Moreover, the proposed solution outperforms a surrogate-based adaptive learning technique and a competing frame selection method. Experiments are also conducted with a recently proposed pre-trained neural network that achieves high classification rates. For this scenario, the FRS-based selection improves the overall classification accuracy from 51.5% to 58.8%.