ASC Model Based on Feature Stratification and Multichannel ECAP A- TDNN

Abstract

The input audio signal in the acoustic scene classification(ASC) task is composed of multiple acoustic events superimposed on each other, leading to problems such as low recognition rate of complex environments and easy overfitting of the model easily. An ASC model based on feature stratification and multichannel ECAPA- TDNN is proposed to address the above problems. Firstly, the extended harmonic-percussive source separation(HPSS) technique is used to divide the log-Mel spectrogram into three components of harmonics, percussive sources and residuals, each of which contains specific types of feature data, to strip the audio signals in the superposition state. On the other hand, the ECAP A-TDNN network structure, which has performed well in the field of acoustic recognition, is applied, and a multichannel ECAP A-TDNN is proposed in combination with the group convolution technique, into which the feature components are input for the ASC task. The results show that the ASC model based on feature stratification can not only reduce the overfitting problem generated by audio overlap, but also enhance the recognition ability of the model for complex environments; moreover, ECAPA-TDNN can achieve a more continuous focus on acoustic features and improve the classification performance while maintaining the original parameter magnitude.