A hierarchical approach for speech-instrumental-song classification

Area of Science:

• Signal Processing
• Machine Learning
• Acoustics

Background:

• Audio classification is crucial for applications like content-based audio retrieval and indexing.
• Existing methods may lack robustness or efficiency in distinguishing complex audio types.

Purpose of the Study:

• To develop a novel, hierarchical scheme for classifying audio signals into three distinct categories: speech, instrumental music, and music with voice (song).
• To improve the accuracy and effectiveness of audio signal classification for various multimedia applications.

Main Methods:

• A hierarchical classification approach was employed.
• The first stage uses audio texture features (Zero Crossing Rate - ZCR, Signal to Total Energy - STE) to differentiate speech from music.
• The second stage utilizes Mel Frequency Cepstral Coefficients (MFCC) and a Random Sample and Consensus (RANSAC) classifier to distinguish between instrumental music and songs.

Main Results:

• The proposed audio texture effectively captures contextual information and summarizes frame-level features.
• The RANSAC classifier demonstrated capability in handling diverse data for music sub-classification.
• Experimental results validated the effectiveness of the proposed hierarchical audio classification scheme.

Conclusions:

• The novel hierarchical scheme provides an effective method for classifying audio signals into speech, instrumental music, and songs.
• The integration of audio texture and MFCC features with a RANSAC classifier offers a robust solution for audio analysis.
• This approach enhances the foundation for advanced audio retrieval and indexing systems.