Speech discrimination in adverse conditions using acoustic knowledge and selectively trained neural networks

Abstract

It is demonstrated that the STNN (selectively trained neural network) method improves confusable work discrimination. Tests conducted on clean and Lombard-noisy speech show that using only a small part (two frames) of the work where useful information for discrimination is located is more efficient than taking into account the whole word. Recognition scores obtained with a continuous-density HMM (hidden Markov model) are lower than those obtained with the proposed method. The present results show an increase in recognition accuracy for the tests on Lombard-noisy speech when the training is done on clean, Lombard, and Lombard-noisy speech. Furthermore, if the same noise is used for the training and the test, the STNN performances improve far more than those of the HMM. The STNN method does not need any precise detection of word boundaries. This influences the robustness of the method, especially in noisy conditions.<>