Environment Sound Classification Using a Two-Stream CNN Based on Decision-Level Fusion.

Area of Science:

• Computer Science
• Artificial Intelligence
• Signal Processing

Background:

• Deep learning models show promise in categorization tasks but struggle with environment sound classification (ESC) using standard features like log-mel spectrogram (LM) and mel frequency cepstral coefficient (MFCC).
• Existing deep neural network models for ESC often yield unsatisfactory performance due to limitations in feature representation.

Purpose of the Study:

• To develop an improved method for environment sound classification (ESC) that overcomes the limitations of current approaches.
• To enhance the comprehensive representation of environmental sounds through novel feature combinations.

Main Methods:

• Proposed two novel combined features for a more comprehensive representation of environment sounds.
• Developed a four-layer convolutional neural network (CNN) to process these aggregated features for ESC.
• Fused CNNs trained on different features using Dempster-Shafer evidence theory to create the TSCNN-DS model.

Main Results:

• The proposed combined features, when used with the four-layer CNN, demonstrated superior performance for environment sound taxonomy.
• The TSCNN-DS model achieved a classification accuracy of 97.2% on the UrbanSound8K dataset.
• This accuracy represents the highest taxonomic accuracy reported for existing models on this dataset.

Conclusions:

• The proposed combined features and CNN architecture are effective for environment sound classification.
• The TSCNN-DS model significantly outperforms conventional methods in ESC tasks.
• The findings suggest a promising direction for improving the robustness and accuracy of environment sound recognition systems.