Problems and solutions for noisy speech recognition

Abstract
Automatic speech recognition has reached high level performances but it usually fails in coping with real-life, noisy environments. An essential reason is the mismatch between the conditions in which a system is trained and used. A large number of solutions have been proposed in order to solve this problem. Those solutions can be classified into two main, non exclusive categories. Firstly, signal processing and parametrization techniques can be used as a preprocessing step in order to enhance the SNR of the corrupted speech signal. Secondly, the different steps of the pattern matching process can be modified in order to account for the effects of noise. This paper presents a brief survey of the noisy speech recognition field. We first summarize the major difficulties that are encountered in the development of a system, and we then introduce three main categories of solutions dealing with acoustical preprocessing and parametrization of the speech signal, statistical modelling, and recognition techniques.