Comparison of several preprocessing techniques for robust speech recognition over both PSN and GSM networks

In this paper several preprocessing techniques used to improve speech recognition performance are compared over both PSN and GSM networks. Recognition experiments are conducted on a digit database in a speaker-independent isolated-word mode in order to evaluate the performances under within- and cross-network (PSN and GSM) conditions. Two classes of preprocessing techniques are distinguished depending on whether they deal with additive ambient noise or convolved perturbations. The first class preprocessing techniques are based on spectral subtraction. In the second class, the low frequencies of cepstral trajectories are eliminated in order to reduce convolved disturbances. Blind equalization adaptive filtering has been proposed to reduce channel effects. In this study, channel equalization and speech enhancement techniques are combined and compared. Different recording conditions may be integrated in order to increase robustness. This is done during the training phase using HMM models with variable parameters. Recognition results are analysed as a function of recording conditions.