Adaptive Noise Dictionary Design For Noise Robust Exemplar Matching Of Speech

This paper investigates an adaptive noise dictionary design approach to achieve an effective and computationally feasible noise modeling for the noise robust exemplar matching (N-REM) framework. N-REM approximates noisy speech segments as a linear combination of multiple length exemplars in a sparse representation (SR) formulation. Compared to the previous SR techniques with a single overcomplete dictionary, N-REM uses smaller dictionaries containing considerably fewer noise exemplars. Hence, the noise exemplars have to be selected with care to accurately model the spectrotemporal content of the actual noise conditions. For this purpose, in a previous work, we introduced a noise exemplar selection stage before performing recognition which extracts noise exemplars from a few noise-only training sequences chosen for each target noisy utterance. In this work, we explore the impact of the several design parameters on the recognition accuracy by evaluating the system performance on the CHIME-2 and AURORA-2 databases.