Room volume classification from room impulse response using statistical pattern recognition and feature selection.

Area of Science:

• Acoustics
• Signal Processing
• Machine Learning

Background:

• Room volume classification from room impulse responses (RIRs) is crucial for acoustic scene analysis.
• Existing methods often require source-to-receiver distance and are sensitive to acoustic absorption.
• There is a need for robust RIR-based room volume estimation methods.

Purpose of the Study:

• To develop and evaluate a novel room volume classification method using RIRs.
• To eliminate the requirement for source-to-receiver distance in RIR-based volume estimation.
• To enhance robustness against variations in room acoustic absorption.

Main Methods:

• Defined novel room volume features extractable from RIRs.
• Employed Gaussian mixture models for room volume class modeling.
• Utilized a maximum likelihood criterion normalized by a background model for classification.
• Performed feature selection using various classification error metrics.

Main Results:

• Achieved an equal error rate of 0.1% for simulated RIRs.
• Attained an equal error rate of 19.1% for measured RIRs.
• Demonstrated the potential robustness of the method to absorption differences.

Conclusions:

• The proposed method effectively classifies room volume from RIRs without source-receiver distance.
• The approach shows promise for acoustic scene analysis applications.
• Further validation with diverse measured RIRs is warranted.