Normalized Autocorrelation based Features for Robust Speech Recognition in Context with Noisy Environment

This paper presents a robust approach for an automatic speech recognition system (ASR) when both additive and convolutional noises corrupt the speech signal. Robust features are derived by assuming that the corrupting noise is stationary and the channel effect is fixed during the utterance. In the proposed method the effect of additive and convolutional distortions are minimized by two stage filtering. The first filtering stage includes differential temporal filtering in the autocorrelation domain for reducing additive noise effects, followed by additional filtering in the logarithmic spectrum domain to reduce convolutional noise effects. Convolutional channel distortion is assumed to be linear and time invariant. A task of multispeaker isolated Hindi word recognition is conducted to demonstrate the effectiveness of using these robust features. The cases of channel filtered speech signal corrupted by white noise and different colored noises such as factory, babble and F16, which are further corrupted by channel distortion are tested. Experimental results show that the proposed method can significantly improve the performance of isolated Hindi word recognition system in noisy environment.