Research On The Key Technologies Of Speech Recognition In Noisy Environment

Noise becomes a major factor which blocks the application of speech recognitionproducts. In order to make the speech recognition products more practical, improving therecognition rate under noisy environment has become an urgent problem to be solved. Basedon the research of related information at home and abroad, several key techniques of speechrecognition in noisy environment are researched in this paper.Pre-processing directly affects the feature extraction and the accuracy of speechrecognition. The detailed analysis of pre-processing of speech signal under the noisyenvironment, speech enhancement (noise removal) in the pre-processing, preemphasis,framing, adding window and endpoint detection are presented in this paper.The separation technology which automatically separates the speech signal from thenoise is researched. The separation technology researched in this paper includes blindseparation algorithm, Independent Component Analysis (ICA) and ICA based on geneticalgorithm. On this basis, blind separation algorithm based on tabu search is proposed. Tabusearch is used for learning method in the searching process, and kurtosis is used as the fitnessof a target optimal separation matrix. The desired speech signal is selected according to thepitch frequency. Experiments show that this method can jump out of local optimization andsearch for global optimization, so that the speech signal and noise signal is separated.Consequently, a more pure speech signal can be obtained to lay the foundation for thefollow-up of speech recognition.The extraction of characteristic parameters with good anti-noise performance is the keytechnology in speech recognition in noisy environment. The Teager-Kaiser Energy operatorhas a good anti-noise performance and Gammatone filter is desirable in characteristics ofhuman hearing. Accordingly, a kind of speech feature parameter TECC is extracted in thisarticle. In view of the fact that differential parameters can reflect the dynamic characteristics,both TECC parameters and differential TECC parameters are used to form thecombined-parameters (TECC+△TECC). Experiments show that TECC parameters have abetter anti-noise and recognition performance in noisy environment than the traditionalspeech parameters LPCC, MFCC and the combined-parameters MFCC+△MFCC.A small vocabulary speech recognition system using MATLAB is realized. The overallimplementation process and its specific operations are also elaborated in this paper.