SRIR Room Acoustics Classification
Comparing PCA, LDA, and EFA for classifying room acoustics from spatial room impulse responses
Statistical comparison of three dimensionality-reduction methods — PCA, LDA, EFA — for classifying rooms by their acoustic character, and for identifying which acoustic parameters actually drive that classification.
Approach
Acoustic parameters (EDT, DT20m, D50, C50/C80, DRR, Grel, TS) were extracted across octave bands (62 Hz–8 kHz) from a high-resolution spatial room impulse response dataset — five rooms measured with a robot-mounted 7-channel spherical microphone array and exponential sine sweeps (Stolz et al., 2024, via Zenodo). Features were z-score standardized, then reduced/classified three ways:
- PCA — unsupervised, variance-based (baseline).
- LDA — supervised, class-separability-based, band-wise and full-matrix.
- EFA — unsupervised, correlation-based latent factor extraction (Kaiser criterion, Varimax rotation).
A stratified 70/30 train/test split fed an LDA-based classifier on each method’s reduced features, with parameter sensitivity checked by excluding parameters sequentially.
Results
| Method | Accuracy |
|---|---|
| PCA | 99.11% |
| EFA | 94.54% |
| LDA | 94.04% |
- PCA gives the best raw classification accuracy but is unsupervised and doesn’t explain why rooms separate.
- LDA identifies DT20m as the most sensitive parameter for classification.
- EFA surfaces DRR and Grel as the critical latent factors, uncovering separate temporal-decay and clarity dimensions that PCA’s variance-only view doesn’t distinguish.
- Confusion matrices show minimal misclassification, concentrated between acoustically similar rooms.
Stack
MATLAB, Statistics and Machine Learning Toolbox, SDMtools.
Status
Complete — full analysis report and presentation included in the repo.