Noise robust digit recognition using sparse representations

Despite the use o f noise robustness techniques, automatic speech recognition (ASR) systems make many more recognition errors than humans, especially in very noisy circumstances. We argue that this inferior recognition performance is largely due to the fact that in ASR speech is typically processed on a frameby-frame basis preventing the redundancy in the speech signal to be optimally exploited. We present a novel non-parametric classification method that can handle missing data while simul­ taneously exploiting the dependencies between the reliable fea­ tures in an entire word. We compare the new method with a state-of-the-art HMM-based speech decoder in which missing data are imputed on a frame-by-frame basis. Both methods are tested on a single digit recognition task (based on AURORA-2 data) using an oracle and an estimated harmonicity mask. We show that at an SNR of -5 dB using the reliable features o f an entire word allows an accuracy o f 91% (using mel-log-energy features in combination with an oracle mask), while a conven­ tional frame-based approach achieves only 61%. Results ob­ tained with the harmonicity mask suggest that this specific mask estimation technique is simply unable to deliver sufficient reli­ able features for acceptable recognition rates at these low SNRs.