Statistical Analysis Of Short-time Speech Signal

The voice signal analysis is the foundation for speech signal processing, becauseit is only possible to analyze the parameters of the speech signal that accuratelyrepresents the essential characteristics, possible to achieve an efficient communicationof voice, speech synthesis, speech recognition processing by the parameters.Moreover, the degree of precision of speech signal analysis also determines the levelof speech recognition and speech synthesis rate sound quality is good or bad, so thatthe voice signal analysis throughout the speech signal processing and applicationwhich has a pivotal position.Although the overall speech signal having a time varying characteristic, but in avery short time zone is relatively stable, i.e., its characteristics remain substantiallyunchanged, this is generally considered to be a short time between the10ms to30ms.Thus, the speech signal in a short time as a smooth process, i.e. a speech signal havinga short-term stationary. Therefore, any voice signal processing and analysis must bebased on a short time, so-called "short-term analysis." Statistical analysis of the text ofthe speech signal is based on the "short-term analysis".The Voiced/Unvoiced Decision is a very important part of voice signalprocessing.The traditional method is to choose a short judgment eigenvalues, and set athreshold limit value of the feature. Although this method is simple, but in fact clear/no clear demarcation between the dullness set threshold that is mandatory for voicingboundaries, which will greatly increase the U/V misjudgment in the overlap region.In order to avoid such drawbacks of conventional methods proposed U/V judgmentmethod based on short-term statistical analysis. The method is based on statisticalanalysis of short-term basis, the first of U/V short priori probability and probabilitydensity function estimate energy distribution, and then with the judgment Bayesianmodel U/V decision. Priori probability estimates is the use of U/V characteristics ofa single threshold, a large number of speech frames into U/V frame, then a prioristatistical probability; probability density function estimation is to use a supervisedparameter estimation method, first draw the U/V histograms, probability densityfunction to determine the type from the histogram, and then estimate the parameters with a large number of voice samples to obtain a probability density function.Simulation results show that the new method has better judgment results thantraditional methods.