Sparse Approximations for Drum Sound Classification

摘要

Up to now, there has only been little work on using features from temporal approximations of signals for audio recognition. Time–frequency tradeoffs are an important issue in signal processing; sparse representations using overcomplete dictionaries may (or may not, depending on the dictionary) have more time–frequency flexibility than standard short-time Fourier transform. Also, the precise temporal structure of signals cannot be captured by spectral-based feature methods. Here, we present a biologically inspired three-step process for audio classification: 1) Efficient atomic functions are learned in an unsupervised manner on mixtures of percussion sounds (drum phrases), optimizing the length as well as the shape of the atoms. 2) An analog spike model is used to sparsely approximate percussion sound signals (bass drum, snare drum, hi-hat). The spike model consists of temporally shifted versions of the learned atomic functions, each having a precise temporal position and amplitude. To obtain the decomposition given a set of atomic functions, matching pursuit is used. 3) Features are extracted from the resulting spike representation of the signal. The classification accuracy of our method using a support vector machine (SVM) in a 3-class database transfer task is 87.8%. Using gammatone functions instead of the learned sparse functions yields an even better classification rate of 97.6%. Testing the features on sounds containing additive white Gaussian noise reveals that sparse approximation features are far more robust to such distortions than our benchmark feature set of timbre descriptor (TD) features.