Speaker dependent isolated word recognition

Recent developments in digital signal processing and computer technologies have made speech signals to be stored, to be compressed, to be synthesized and to be recognized digitally possible. In speaker dependent isolated word recognition implementation, two methods have been utilized in feature extraction of speech signals: Linear Predictive Coding, LPC, and Mel-Frequency Cepstrum Coefficients, MFCC. The results by LPC and MFCC are compared. After the speech signals are converted from analog to digital and stored in appropriate format, the basic parameters of speech features are extracted. The minimum distance between feature parameter vectors of the template words and the feature parameter vectors of word to be recognized is calculated by Dynamic Time Warping, DTW, method. Speaker dependent isolated word recognition is implemented on computer based system. In the training phase, a user records the words in computer. In word recognition phase, system tries to recognize which word is uttered by the user. Matlab technical computation programming language is used for this study. In order to measure the performance of system, isolated words are chosen the numbers between 0 and 9 from Turkish and English language. English numbers are recognized by 85-95 % and Turkish numbers are recognized by 75-90 %. In general, MFCC method has much better performance than LPC method to recoginize isolated words. An appropriate method has been investigated in order to eleminate the negative effect of background noise on the word recoginition. In denoising the speech signal, Discrete Wavelet Tranformation, DWT, is utilized. It is observed that with the combined use of DWT in recording, the performance of Turkish isolated word recognition is very near to the performance of English isolated word recognition.Key Words : Word recognition, Linear Predictive Coding, Mel-Frequency Cepstrum Coefficients