General audio tagging with ensembling convolutional neural networks and statistical features.

Area of Science:

• Computer Science
• Machine Learning
• Signal Processing

Background:

• Audio tagging, the process of assigning descriptive labels to audio clips, faces challenges due to limited data and inherent label noise.
• Existing methods struggle to effectively leverage diverse data types and mitigate inaccuracies in training datasets.

Purpose of the Study:

• To develop a robust audio tagging system that overcomes data limitations and label noise.
• To enhance the accuracy and reliability of audio event detection and classification.

Main Methods:

• An ensemble learning framework was designed to integrate statistical audio features with outputs from deep learning classifiers.
• A novel sample re-weighting strategy was implemented within the framework to address the issue of noisy labels.
• The approach combined complementary information from diverse feature sets and models.

Main Results:

• The proposed method achieved a mean average precision (mAP) of 0.958.
• This performance significantly outperformed the baseline system, demonstrating the effectiveness of the ensemble and re-weighting strategies.
• The integration of statistical and deep features proved beneficial for audio tagging.

Conclusions:

• The developed ensemble learning framework with sample re-weighting offers a superior solution for audio tagging.
• The findings highlight the importance of addressing noisy labels and utilizing complementary information for improved audio analysis.
• This approach provides a promising direction for advancing the field of audio event detection.