Multi-band CNN architecture using adaptive frequency filter for acoustic event classification

Although Convolutional Neural Networks (CNNs) architecture based learning systems have shown impressive results in the performance of numerous classification tasks, their effectiveness has been limited in certain cases of acoustic based classification. This vulnerability is particularly evident in the acoustic event classification tasks using spectral features. For example, spectral based features may suffer from a typical normalization process when it is fed to a neural network for training since the magnitudes in high-frequency band are inadvertently attenuated even though they may yet contain useful discriminant features. Although some research efforts try to mitigate this problem by introducing a multi-band approach for attaining salient and stable features, it requires empirically preset frequency bands to separate the spectral features. Being heuristic, however, this process is difficult to ensure the consistency required for high correlation between manually separated features and good classification performance. In this paper, we propose a novel filter parameter modeling framework performing optimized frequency sub-band separation via CNN based end-to-end training for achieving high acoustic event classification performance. In particular, the filter response characteristics, namely, cut-off frequencies and damping ratio for roll off are considered as added learning parameters to the CNN architecture for the proposed end-to-end learning framework so that the filter's frequency response is optimized for producing salient features. The proposed training process is shown to not only automatically select the filter parameters for multi-band frequency separation but also guarantee high correlation between the resulting sub-band features and accurate classification performance.