Robust speech recognition under noisy environments using Lombard effect compensation and dynamic characteristic

Abstract

Although speech recognition systems in artificially constrained conditions have already reached high levels of performance, they degrade dramatically when they are applied in the real world, particularly in noisy environments. In noisy environments human speech productions are influenced by noises (Lombard effect) and speech signals are contaminated in ways which affect the performance of speech recognition systems. This thesis describes a Lombard effect compensation and noise suppression method of improving speech recognition performance in noisy environments. First of all, to become familiar with the characteristics of speech affected by the Lombard effect, several features of Lombard speech were investigated. We found that vocal intensity was varied by the Lombard effect, and that the Lombard effect had a statistically significant influence on energy below 500Hz low-band spectral tilt, and the center of gravity determining the spectral structure of speech. Since speech production variations due to the Lombard effect depends on the intensity of the Lombard effect, we formulated a quantitative measure of the Lombard effect level so as to model it more explicitly. Statistical tests were used to discover those features that represent the Lombard effect. A speech degradation model is proposed in order to characterize the distortions of speech in noise and under the Lombard effect. Variations in formant location, formant bandwidth, pitch, spectral tilt, and energy in each frequency band under the Lombard effect are represented by frequency warping and amplitude scaling of each frequency band. Another Lombard effect, the variation of vocal intensity is modeled by a multiplication term depending on the energy of the input speech. Noise contamination is represented by an additive term in the frequency domain. The distortions of noisy Lombard speech are then canceled out according to the speech degradation model. First, spectral subtraction is used to suppress the...