Robustness Improvement of an Automatic Sounds Recognition System by HMM Adaptation to Real World Background Noise

Summary form only given. This work forms part of a larger investigation into the integration of sound surveillance in a monitoring application. However, mismatches between training and testing environment severely degrade performance. Thus, in order to enhance the system robustness, we explored two issues: the training mode and the model adaptation. First, the originality of our system resides in the HMM training mode which consists in using both clean and noisy sets. Our paper proposes a multi-style training approach: the training database includes different levels of real world background noises and the recognizer can be successfully tested in every noisy environment The second robustness improvement procedure is applying environmental adaptation techniques to the baseline recognizer. The algorithms closely examined are maximum likelihood linear regression (MLLR), maximum a posteriori (MAP) and the MAP/MLLR algorithm that combines MAP and MLLR. Experimental evaluation on environmental adaptation using MAP, MLLR and MAP/MLLR techniques illustrates a recognition improvement over the baseline system (i.e. none adapted EI system) results