Toward a Practical Implementation of Exemplar-Based Noise Robust ASR

In previous work it was shown that, at least in principle, an exemplar-based approach to noise robust ASR is possible. The method, sparse representation based classification (SC), works by modelling noisy speech as a sparse linear combination of speech and noise exemplars. After recovering the sparsest possible linear combination of labelled exemplars, noise robust posterior likelihoods are estimated by using the weights of the exemplars as evidence of the state labels underlying exemplars. Although promising recognition accuracies at low SNRs were obtained, the method was impractical due to its slow execution speed. Moreover, the performance was not as good on noisy speech corrupted by noise types not represented by the noise exemplars. The importance of sparsity was poorly understood, and the influence of the size of the exemplar-dictionary was unclear. In this paper we investigate all these issues, and we show for example that speedups of a factor 28 can be obtained by using modern GPUs, bringing its execution speed within range to practical applications.

speech recognition

• speech recognition