Mean absorption estimation from room impulse responses using virtually supervised learning.

Area of Science:

• Building acoustics
• Acoustic signal processing
• Machine learning in acoustics

Background:

• Estimating acoustic properties of existing rooms is crucial for building acoustics.
• Traditional methods often rely on the diffuse sound field hypothesis, which has limitations.
• Room impulse response (RIR) contains rich information about a room's acoustic characteristics.

Purpose of the Study:

• To develop and investigate a novel approach for estimating mean absorption coefficients directly from RIRs.
• To utilize virtually supervised learning and artificial neural networks for this inverse problem.
• To compare the performance of the proposed method against classical formulas.

Main Methods:

• Employed virtually supervised learning with artificial neural networks (ANNs) to map RIRs to absorption coefficients.
• Focused on simple, well-understood ANN architectures.
• Trained models on a simulated dataset with parameters relevant to building acoustics.
• Compared ANN-based estimations with classical formulas requiring room geometry and reverberation times.

Main Results:

• ANN models demonstrated the ability to estimate mean absorption coefficients solely from RIRs.
• The proposed approach overcomes limitations of classical formulas, particularly those related to the diffuse sound field hypothesis.
• Performance on real RIRs at 1 kHz and above is comparable to classical models when reverberation times are reliably estimated.
• The method remains effective even when reverberation times cannot be reliably determined.

Conclusions:

• The developed ANN-based approach offers a robust alternative for estimating mean absorption coefficients in building acoustics.
• This method expands the applicability of acoustic diagnosis, especially in scenarios where traditional assumptions are violated.
• The findings suggest a promising direction for acoustic analysis using machine learning on RIR data.