Table 2. The twelve combinations of feature-type and feature-summarisation tested.
The feature-type and feature-summarisation method jointly determine the dimensionality of the data input to the classifier.
| Label | Features | Summarisation | Dimension |
|---|---|---|---|
| mfcc-ms | MFCCs (+deltas) | Mean & stdev | 52 |
| mfcc-maxp | MFCCs (+deltas) | Max | 26 |
| mfcc-modul | MFCCs (+deltas) | Modulation coeffs | 260 |
| melspec-ms | Mel spectra | Mean & stdev | 80 |
| melspec-maxp | Mel spectra | Max | 40 |
| melspec-modul | Mel spectra | Modulation coeffs | 400 |
| melspec-kfl1-ms | Learned features, 1 frame | Mean & stdev | 1,000 |
| melspec-kfl2-ms | Learned features, 2 frames | Mean & stdev | 1,000 |
| melspec-kfl3-ms | Learned features, 3 frames | Mean & stdev | 1,000 |
| melspec-kfl4-ms | Learned features, 4 frames | Mean & stdev | 1,000 |
| melspec-kfl8-ms | Learned features, 8 frames | Mean & stdev | 1,000 |
| melspec-kfl4pl8kfl4-ms | Learned features, 4 frames, two-layer | Mean & stdev | 1,000 |