A Comparative Survey of Feature Extraction and Machine Learning Methods in Diverse Acoustic Environments.

Area of Science:

• Computational Acoustics
• Machine Learning Applications
• Signal Processing

Background:

• Acoustic event detection and analysis are crucial for applications like elderly monitoring, surveillance, multimedia retrieval, and biodiversity assessment.
• Effective sound source identification is a key technological challenge in these diverse applications.
• Varied sound types and environments necessitate robust artificial intelligence algorithms.

Purpose of the Study:

• To conduct a comparative study of feature extraction and machine learning algorithms for sound source identification.
• To identify the optimal combination of algorithms and feature extraction methods for general-purpose acoustic analysis.
• To evaluate the reliability of proposed methods across different acoustic environments.

Main Methods:

• Feature extraction techniques: Mel Frequency Cepstrum Coefficients (MFCC), Gammatone Cepstrum Coefficients (GTCC), and Narrow Band (NB).
• Machine learning algorithms: k-Nearest Neighbor (kNN), Neural Networks (NN), and Gaussian Mixture Model (GMM).
• Comparative testing across five distinct acoustic environments.

Main Results:

• Most feature extraction and machine learning combinations yielded acceptable results in various acoustic corpora.
• The combination of Gammatone Cepstrum Coefficients (GTCC) with k-Nearest Neighbor (kNN) significantly outperformed other methods.
• Detailed analysis of the GTCC-kNN combination's performance across all tested corpora.

Conclusions:

• The GTCC feature extraction method paired with the kNN algorithm represents a best-practice approach for sound source identification.
• This combination offers high reliability and generalizability across diverse acoustic environments.
• Further analysis confirms the superiority of GTCC-kNN for robust acoustic event detection.