Evaluation of a feature selection scheme on ICA-based filter-bank for speech recognition

In this paper, we propose a new feature selection scheme that can contribute to an ICA-based feature extraction block for speech recognition. The initial set of speech basis functions obtained in Independent Component Analysis (ICA) training phase, has some redundancies. Thus, finding a minimal-size optimal subset of these basis functions is rather vital. On the contrary to the previous works that used reordering methods on all the frequency bands, we have introduced an algorithm that finds optimal basis functions in each discriminative frequency band. This leads to an appropriate coverage of various frequency components and easy extension to other data is also provided. Our experiments show that the proposed method is very useful, specifically in larger vocabulary size tasks, where the selected basis functions trained using a limited dataset, may get localized in certain frequency bands and not appropriately generalized to residual dataset. The proposed algorithm surmounts this problem by a local reordering method in which contribution of a basis function is specified with three factors: class separability power, energy and central frequency. The experiments on a Persian continuous speech corpus indicated that the proposed method has led to 17% improvement in noisy condition recognition rate in comparison to a conventional MFCC-based system.