Generalized Filter-bank Features for Robust Speech Recognition Against Reverberation

Traditionally, automatic speech recognition (ASR) uses a Hidden Markov Model with Gaussian Mixture Model (HMM-GMM) as acoustic model and hand-designed features such as Mel-frequency Cepstral Coefficient (MFCC) as acoustic features. It is usually assumed that the features are uncorrelated, making it possible to use diagonal covariances for the GMM. The assumption generally holds due to the use of Discrete Cosine Transformation (DCT) that de-correlates the speech spectra. However, DCT could cause some information loss, such as correlations between the feature components. Current ASR systems, which is based on Deep Neural Network (DNN) show to be better especially in reverberant conditions when more primitive features, such as filter-bank (FBANK), are used. This might be because DNN is better in modeling non-linear relations between the components of the features. But the use of short-time processing in FBANK may cause the lost of long-term correlations in a speech pattern. To tackle this, we propose a new feature, q-FBANK, which is a generalization of FBANK. The results on artificially reverberant speech show that the proposed features achieve better performance than MFCC and FBANK on DNN-HMM systems where an average error reduction up to 39.73% and 13.5% were achieved respectively.