Research on Dog Bark Recognition Algorithm Based on Mel-Frequency Cepstral Coefficients and Lightweight Neural Network

Abstract

Dog barks serve as key bioacoustic signals for pet monitoring and urban noise early warning. Conventional recognition methods suffer severe feature distortion in noisy environments, while deep learning models carry excessive parameters and slow edge inference. This work proposes an algorithm combining spectral entropy-based adaptive denoising MFCC and ECA lightweight attention network. It fuses static and temporal differential cepstral features and uses audio augmentation to mitigate small-sample overfitting. Tests on ESC50, UrbanSound8K and self-built dog bark datasets show 4.3% higher accuracy than standard MFCC at 0 dB SNR. The model has only 2.07 M parameters and 0.31 G computation load, cutting inference latency by 62.8% versus CNN-LSTM, with an average F1 of 0.912 (5–20 dB SNR). Edge tests validate its real-time performance for embedded bioacoustic detection.