Analysis And Research On Speech Characteristics Based On Non-reconstructed Compressed Sensing

In recent years, with the rapid development of modern information technology, the traditionalNyquist sampling already can’t satisfy people’s growing demand information, and compressedsensing theory，because of its sampling rate is far lower than the Nyquist sampling, thesampling method is simple, much attention has been paid as it caused a huge response ofacademia. In this thesis, the speech signal processing technology is combined with compressedsensing technology, analysis and research on speech characteristics based on non-reconstructedcompressed sensing is discussed.This thesis researches new features of speech compressionsampling observation sequence, thus extract speech feature parameters directly from theobservation sequence, such as formant and pitch.First of all, this thesis studies the speech endpoint detection method based on higher-ordercumulant of the observation sequence, this thesis expounds the theory of higher order cumulant,and analyzes its features, namely for gaussian process, the third order and the higher-ordercumulant above is zero. Then based on compressed sensing theory, analyzes the speechobservation sequence under the gaussian random observation matrix, it is concluded that thespeech observation sequence is not gaussian,and the noise observation sequence is gaussian,based on this, the higher-order cumulant theory is applied to the speech endpoint detection basedon the observation sequence. For unvoiced speech frame and noise frame is difficult todistinguish, this thesis use the rate of full-frequency band energy and low-frequency bandenergy as the second parameter. This article compared the method with the endpoint detectionmethod based on cepstrum distance, this method has better robustness.Second, the thesis analysis the waveform characteristics of speech observation sequence, anddoes some research on the relation between the period of observation sequence and the period ofthe original speech. Then wavelet transform is used for the observation sequence, due to pitchinformation is low frequency information, so the thesis uses low frequency wavelet coefficientsto do autocorrelation, then the pitch period is deteced. Aiming at the condition of the originalspeech with noise, this thesis first do wavelet threshold denoising to the observation sequence,then pitch detection is used, the experimental results show that after wavelet threshold denoising,the accuracy of pitch detection is higher and the noise’s influence on the accuracy of detection isreduced. Third, the thesis deduced power spectrum estimation method from observation sequence,through the estimated original speech autocorrelation sequence, the power spectrum can besolved. Based on strict mathematical proof, this thesis use the autocorrelation matrix ofobservation sequence and the autocorrelation and cross-correlation between the row vector ofobservation matrix as the estimate of the original signal autocorrelation data, successfullyestimate the original signal autocorrelation sequence, and then the power spectrum of theoriginal signal. Because this method does not use iterative reconstruction algorithm, thus thesignal sparsity is not so hihgly required compared to other power spectrum estimation algorithmsusing iterative reconstruction algorithm, so the unvoiced frame is also well estimated.