Analysis And Research On Speech Characteristics Based On Compressed Sensing

With the rapid development of information technology in recent years, the sampling rate limitof the traditional Nyquist sampling theory has not kept pace with people‘s demand for increasingamount of information, so compressed sensing becomes a research focus with the characteristic ofsampling rate lower than that of Nyquist at home and abroad. This paper combines compressedsensing technology and speech signal processing to do some research on characteristics ofcompressed sensing observation sequence, and this paper proposes the method of voicing-stateidentification, detection for pitch period and speech endpoint.First, based on the theory of compressed sensing, this paper discusses the characteristic ofspeech signal observation sequence after compressed sensing, and does some research on theapplicability of the energy and zero crossing rate discrimination of traditional Nyquist sampling incompressed sensing. Then this paper analyzes the voicing-state characteristics based on the speechdigital model, and can draw a conclusion that the unvoiced observation sequence has thecharacteristic of Gaussian signal and the voiced observation sequence has the characteristic ofnon-Gaussian signal. According to the characteristic, this paper designs a voicing-stateidentification algorithm of third-order accumulation directly based on observation sequence, andcompares it with the energy discrimination method of traditional Nyquist sampling in accuracy andcomputing.Second, this paper analyzes the characteristics of observation sequence after compressedsensing, and does some research on the relation between the period of observation sequence and theperiod of the original speech. Based on auto correlation function of observation sequence, this papercan get the period of original speech, and the observation sequence of unvoiced does not present thecharacteristic, which is the basis of voicing-state identification and period detection.Third, this paper analyzes the basis of voice energy estimate based on the observationsequence of compressing sensing. The estimated voice energy is as the speech endpoint detection,and this paper makes the speech endpoint detection with the pink noise, Gaussian noise and carnoise, and makes the comparison with the cepstral distance based on speech observation sequenceand the energy endpoint detection based on traditional Nyquist sampling, reducing the amount ofcomputation. Fourth, power spectrum density of the original speech signal can be derived from the realationbetween the original speech signal and the observation sequence of compressing sensing. Powerspectrum density of the original speech signal can be sparse and can be combined with compressingsensing, so it can be estimated by signal reconstruction based on compressing sensing.And theoriginal speech signal also can got from the signal with noise by spectral subtraction. This papermakes the comparison between the estimated power spectrum density from the observationsequence of compressing sensing and makes a brief analysis.