The Speech Signal With Noise Pitch Detection Technology Research

With the rapid development of the information technology, voice technology isbeing widely used with increasing frequency to many areas. Analyzing speech signalprecisely and then extracting the characteristic parameters that characterize the essenceof speech is an important part of speech signal processing. Pitch period is a veryimportant parameter to describe the excitation source, and extracting pitch periodaccurately is of great significance to various field of speech signal processing such asspeech synthesis, speech coding, speech recognition, speaker recognition, speechsegregation and so on. The commonly used pitch detection methods have goodperformance for clean speech, but the effectiveness of these methods decreasesdrastically even unable to detect with low SNR. However, in practical, speech isinevitably influenced by background noise, so this paper focuses on the pitch detectionfor the noisy speech, and the objective is to develop algorithms with better accuracy androbustness. Based on the research and study of the predecessors, the main works of thispaper are as follows:1. The advantages and disadvantages of the commonly used pitch detectionmethods are introduced and the basic properties which related to the pitch detection areanalyzed and simulated.2. The preprocess algorithm that based on enhancement of frequency domain andimproved band-partitioning spectral entropy is studied. In order to reduce the impact ofnoise, the improved spectral subtraction to enhance the speech in frequency domain isapplied, and the method of improved weighted band-partitioning spectral entropy isproposed to decide the voiced speech and unvoiced speech. Experimental resultsillustrate that the proposed preprocess algorithm reaches a better performance.3. In this paper we present a better anti-noise algorithm which uses a circularaverage magnitude sum function (CAMSF) weighted by the inverse of circular averagemagnitude difference function (CAMDF) with reference of the weighted autocorrelation(WAUTOC) method. Major innovations include: utilizing the circular averagemagnitude difference function (CAMDF) to replace the average magnitude difference function (AMDF); establishing the circular average magnitude sum function (CAMSF)instead of the autocorrelation function (ACF); improving the median smoothing postprocessing.4. A pitch period detection based on harmonic sinusoidal autocorrelation model(HSAC) and time-domain matching scheme is presented by combining time domain andfrequency domain analysis. Employing the HSAC model and least-squares fittingoptimization technique is develop to extract a pitch harmonic from the noisy speech.Then the harmonic number associated with the pitch harmonic is determined through atime-domain matching scheme of CAMSF weighted impulse-train. Last a pitch trackingscheme using dynamic programming is devised to obtain a smoothed pitch contour. Bysimulation experiments, it is shown that the proposed pitch extraction method is usefulin noisy environments.