雙變量時間數列的統計頻譜估計與區別分析-以兩個人說中文的聲音辨識為例 羅琪,張效豪
應用統計學系 管理學院 [email protected]
摘要
The extension of classical pattern-recognition techniques to
experimental time series data is a problem of great practical interest. An important application in engineering is to the problem of discriminating between various speech patterns. Throughout the engineering literature, most approaches assumed very specific Gaussian additive signal and noise models and then developed the discriminant criteria to minimize errors. In general, this requires that we assume prior knowledge of the signal
waveforms and spectra under each of the hypotheses, so that discriminant functions can be calculated for an observed time series. In this
dissertation, the spectra of two speakers on speaking Chinese words “open the door, sesame” are assumed to be unknown. The estimation and
hypothesis-testing problems are formulated in terms of sample spectral densities with sample approximate distributions. Finally, frequency domain approximations are used to the optimum discriminant functions to identify the speech patterns of two speakers on speaking Chinese words “open the door, sesame”. Since each person has left and right two series (bivariate time series) in one record, the estimation of cross spectrum, phase
spectrum and coherency between two series are also considered. The
estimated quadratic discriminant function is quite well in classifying the two speakers’ voices because the average apparent error rate of 45 pairs of 10 speakers is only about 1.528%. This indicates that it is a good way to apply the frequency domain discriminant analysis of time series to do the speech recognition on Chinese speaking of two speakers.
關鍵字：bivariate time series, spectral estimation, discriminant analysis, speech recognition
