Research For Time-frequency Features Extraction Of Chinese Initials Based On MP Algorithms

Speech is the most used, the most important information carrier when human beings tocommunicate with each other. Chinese syllable is composed of initials, finals and tones,structure of the initials and finals is unique to the Chinese syllable. Chinese initials are allconsonants, non-stationary features of Chinese initials include incentives similar to Gaussianwhite noise, waveform change dramatically, easy to noise pollution. The time-frequencyanalysis method is powerful tool for non-stationary signal, time-frequency resolution of theShort Time Fourier Transform can not be adaptive change, the wigner-ville distribution existcross-interference term defects. In this paper, MP sparse decomposition is used for Chineseinitials atomic parametric time-frequency feature extraction, specific research are as follows:Firstly, the time-frequency feature extraction of Chinese initials based on Chirp atoms isstudied. The over-complete Chirp time-frequency atom library is used, sparse representationof Chinese initials time-frequency distribution is consisting of MP decomposition atomicPWVD accumulation, iterative threshold is90%of the initial energy, the time-frequencyreconstruction of the MP sparse decomposition Chirp atoms effective inhibitcross-interference terms, compared with Gabor atoms, this method was more refinedportrayed initials time-frequency energy distribution.Secondly, the time-frequency feature extraction of Chinese initials based on vocalmechanism is studied. Due to the different of vocal method and vocalization parts, Chineseinitials are divided into voiced consonants, plosive, affricates, fricatives four categories.Morlet wavelet, Gabor and Chirp atom three time-frequency dictionary is used for fourcategories initials MP time-frequency feature extraction. Simulation results show that: forharmonic voiced consonants, Gabor atoms using the number of atoms and lesstime-consuming with outstanding performance; for affricates and fricatives, number of Chirpatoms is90%of the Gabor atom, residual energy decay rate faster than Gabor atoms.Finally, the time-frequency feature extraction of Chinese initials based on human earperception is studied. Against the specific situation which Chinese initials incentives similarto the Gaussian white noise, are highly susceptible to noise pollution, Semi-ERB scaleGammatone filter group human ear perceives model is selected, the process of the prior modelis improved, including band scale selection and Gammatone filter bank parameters design, inthe signal-to-noise ratio of40dB Gaussian white noise environment,after filtering the signal isused to time-frequency feature extraction studies based on atomic MP sparse decomposition.Simulation results show that: the two kinds of atoms with better reconstruct the original signalwaveform in40dB noise environment; for Plosive p, Gabor atomic number less than Chirpatoms in the first frame, from the three decomposition overall data, number of Chirp atoms is94.74%of the Gabor atom; for voiced consonants r, the Gabor atoms use the number and timeless.