Research on the Evaluation Model of Sound Quality in Vehicles Based on Dynamic Activated Mel-Spectrogram

To address the high experimental cost of the subjective evaluation of interior vehicle sound quality, this paper proposes an objective evaluation model of sound quality based on the annoyance level of interior noise. First, noise samples of different models under different working conditions were collected. Second, subjective experiments were carried out with annoyance as the evaluation index to construct the in-vehicle noise data set. In order to include both static and continuity features in the model input, we performed two differencing and activation of the Mel-Spectrogram to extract a new dynamic activated Mel-Spectrogram (DAM) by using the original Mel-Spectrogram to learn the dynamic weights obtained after activation. Then the DAM is fed into ResNet152 (Residual Networks 152) for sound quality prediction and the network is optimized using ECA (Efficient Channel Attention). After a large amount of data training, the model obtained an accuracy of 98.87% on the test set. Finally, according to the classification accuracy and time consumed, the proposed model is compared with other models, and the comparison results show that the proposed model has excellent performance and good sound quality evaluation ability, which can lay a practical foundation for sound quality improvement tasks.