Environmental noise present in real-life applications substantially degrades the performance of speech recognition systems. An example is an in-car scenario where a speech recognition system has to support the man-machine interface. Several sources of noise coming from the engine, wipers, wheels etc., interact with speech. Special challenge is given in an open window scenario, where noise of traffic, park noise, etc., has to be regarded. The main goal of this thesis is to improve the performance of a speech recognition system based on a state-of-the-art hidden Markov model (HMM) using noise reduction methods. The performance is measured with respect to word error rate and with the method of mutual information. The noise reduction methods are based on weighting rules. Least-squares weighting rules in the frequency domain have been developed to enable a continuous development based on the existing system and also to guarantee its low complexity and footprint for applications in embedded devices. The weighting rule parameters are optimized employing a multidimensional optimization task method of Monte Carlo followed by a compass search method. Root compression and cepstral smoothing methods have also been implemented to boost the recognition performance. The additional complexity and memory requirements of the proposed system are minimum. The performance of the proposed system was compared to the European Telecommunications Standards Institute (ETSI) standardized system. The proposed system outperforms the ETSI system by up to 8.6 % relative increase in word accuracy and achieves up to 35.1 % relative increase in word accuracy compared to the existing baseline system on the ETSI Aurora 3 German task. A relative increase of up to 18 % in word accuracy over the existing baseline system is also obtained from the proposed weighting rules on large vocabulary databases. An entropy-based feature vector analysis method has also been developed to assess the quality of feature vectors. The entropy estimation is based on the histogram approach. The method has the advantage to objectively asses the feature vector quality regardless of the acoustic modeling assumption used in the speech recognition system.    «

Environmental noise present in real-life applications substantially degrades the performance of speech recognition systems. An example is an in-car scenario where a speech recognition system has to support the man-machine interface. Several sources of noise coming from the engine, wipers, wheels etc., interact with speech. Special challenge is given in an open window scenario, where noise of traffic, park noise, etc., has to be regarded. The main goal of this thesis is to improve the performance...    »