NOISE-ROBUST DIGIT RECOGNITION WITH EXEMPLAR-BASED SPARSE REPRESENTATIONS OF VARIABLE LENGTH

This paper introduces an exemplar-based noise-robust digit recognition system in which noisy speech is modeled as a sparse linear combination of clean speech and noise exemplars. Exemplars are rigid long speech units of different lengths, i.e. no warping mechanism is used for exemplar matching to avoid poor time alignments that would otherwise be provoked by the noise and the natural duration distribution of each unit in the training data is preserved. Speech and noise separation is performed by applying non-negative sparse coding using a separate exemplar dictionary for each labeled unit (in this case half-digits) rather than a single dictionary of all units. This approach does not only provide better classification of speech units but also models the temporal structure of speech and noise more accurately. The system performance is evaluated on the AURORA-2 database. The results show that the proposed system performs significantly better than a comparable system using a single dictionary at positive SNR levels.