An Evaluation Method for Interpretable Models in Sound Recognition

Abstract In recent years, the application scope of artificial intelligence voice recognition has been expanding rapidly. Investigating the interpretability of sound recognition and enhancing the effectiveness of sound interpretation hold significant importance for both theoretical research and practical applications. However, there is currently a lack of an effective method to assess the interpretability of sound explanations. Therefore, this paper introduces an evaluation method called AI-SCORE, which focuses on interpretable models for sound recognition. AI-SCORE processes the features of sound spectrograms based on their interpretability importance and then re-predicts to calculate the model's positive and negative feature correlations. Based on this, we examine the impact of different feature extraction methods, various neural network models, and parameter settings during the interpretation on interpretation performance. Experiments show that under the ResNet network model, the model trained using mel spectrogram as a feature extraction method achieves the best interpretation performance. Additionally, appropriate iterative precision can further enhance the interpretation effect during the interpretation process.