Polynomial Regression And Adaptation For Noise Robust Recognition

Prevailing speech recognition system could obtain a good recognition accuracy forclean speech data, but their performance will degrade rapidly in noisy environmentsowing to the mismatch between the acoustic models and the testing speech. There-fore, noise robust speech recognition as a hot issue receives wide attention and hasbeen extensively studied.However, these researches can not fully meet the practicalrequirement due to the complexity of the noisy environments. This thesis investigatesa novel and more effcient extension of GVP-HMM that can also model the trajectoriesof feature space linear transforms. The transforms are trained under the constrainedmaximum likelihood linear regression (CMLLR) criterion and would be applied onfeatures with auxiliary information to eliminate the mismatch from clean model andtest environment.In this thesis, the theories of GVP-HMM and derivation of the feature space trans-forms are introduced. In both of the algorithms above, polynomial regression functionare used to present the variable parameter trajectories. We also describe the systemstructure of feature GVP-HMM in training and test stage. The main advantage of us-ing variable parameters in feature domain is avoidance of complex computation andadjustment in model space. The results of experiment shows that extension of thevariable parameter on feature spaces could achieve a signifcant error rate reductionsover the multi-style training baseline system, and computational cost when transformsapplying. Based on modelling the parameters trajectories on model and feature spacesimultaneously, a hierarchical system that cascade different types of modelling param-eters is proposed to compensate the mismatch both in model and feature domain. Byusing two sides of variable parameters that vary with auxiliary SNR, this hierarchicalsystem shows a good performance on noisy robust speech recognition.This thesis also investigates a robust speech recognition method that combinesGVP-HMM systems and speaker adaptation. In this method, different types of variableparameters that modeled with auxiliary noise condition are used to form the basicacoustic model. And then, speaker adaptation transforms are generated on that “anti-noise” model. This system makes full use of auxiliary noise and speaker information to adjust model parameters and shows powerful robustness performance.Three recognition tasks were set up to evaluate those methods above, and eachof them showed better performance over their baselines systems. The results demon-strated that the extension of GVP-HMM to feature space could achieve recognitionaccuracy increment with lower storage space and small computational cost. Experi-ment also proved that the combination system of GVP-HMM and speaker adaptationcould effectively overcome the two hot problems in robust speech recognition.