Acoustic Scene Classification by Linear Projections for Dimensionality Reduction.

We present an acoustic scene classification method with low storage requirements. It employs dimensionality reduction techniques based on linear projections of the input data. We evaluated our method with the DCASE 2020 Task 1B dataset. The audio signals were described via image texture descriptors such as LBP, LPQ, and spectral features. We showed that it is possible to project an early fusion of all 387 features into only 32 dimensions using PCA and NMF with only a slight loss in classification performance. The best result, respecting the 500KB size limit imposed on the challenge, was obtained with the NMF approach and 16 dimensions. In this case, the generated model has been reduced by about 95.8 percentage points concerning the original size with 387 attributes. The F1-score obtained is 0.87, and the model size is 393. 5KB.