Least squares filtering of speech signals for robust ASR

The behavior of the least square filter (LeSF) is analyzed for a class of non-stationary signals that are composed of multiple sinusoids whose frequencies, phases and the amplitudes may vary from block to block and which are embedded in white noise. Analytic expressions for the weights and the output of the LeSF are derived as a function of the block length and the signal SNR computed over the corresponding block. Recognizing that such a sinusoidal model is a valid approximation to the speech signals, we have used LeSF filter estimated on each block to enhance the speech signals embedded in white noise. Automatic speech recognition (ASR) experiments on a connected numbers task, OGI Numbers95 [20] show that the proposed LeSF based features yield an increase in speech recognition performance in various non-stationary noise conditions when compared directly to the un-enhanced speech and noise robust JRASTA-PLP features.