| 摘要 |
FIELD: physics; acoustics. ^ SUBSTANCE: invention concerns technical solutions used with reference to vehicles "black boxes" for allocation of spectral characteristics of sounds of speech. According to the method, an ultrasonic signal is registered, its transformation to a digital form is carried out, the digitised signal is broken into overlapped windows with bias of a reference mark of each subsequent window concerning by the beginnings of each previous window. Signal fragments in each window are exposed to discrete Fourier transformation and the bank of the obtained real and imaginary parts is created. The Fourier spectrum is calculated and normalised, basic signals are shaped in the form of wavelet-functions and their resultants are obtained with entourage of every component of the normalised Fourier spectrum, the obtained resultants are consolidated, forming a resultant matrix of resultants. Boundary conditions for band segmentation of a general voice component are formed by splitting to subbands. The components of a resultant matrix of resultants are separated for the frequency range of a general voice component, having coefficients, multiple to frequency of a general component, and they are summarised. The peak value of a resultant matrix and argument corresponding to the peak value is defined for each of subband frequencies of a general voice component, the vector of informative signs representing range of pairs of maximum values and arguments corresponding to them are formed. Standards of similarity of pairs are formed with use of range of vectors of informative signs and standards of similarity of pairs, weights of sequence of the informative signs, characterising presence of a lined spectrum in a signal. By means of the generated weights of sequence of informative signs sequences of the components of informative signs are sorted out and the spectrums possessing flatness and smoothness of general component frequency dynamics are separated; real and imaginary parts of the separated spectrums are chosen from the bank of real and imaginary parts of Fourier transformation and registered in the form of amplitude-frequency characteristics of voiced sounds. The method is realised by the system containing consistently joined numeral recorder, the digitisation block, the block of discrete Fourier transformation, a block of Fourier spectrum normalisation, a shaper of a resultant of a matrix of convolutions, a adder, a registrar of the peak values, the shaper of a vector of signs, a block of lag lines, a block of formation of weights of sequence of informative signs, a block of search of sequence a builder of informative signs and allocation of the spectrums possessing flatness and smoothness of dynamics of frequency of a general component, a comparator, a selector of components of Fourier transformation and the block of recording of informative signs. Also the power unit, a storage block, a generator of basic signals, a shaper of frequency band parametres dissection of a voice general component, a shaper of standards of similarity and a shaper of threshold level are included into system. ^ EFFECT: increase of accuracy of definition of parametres of voiced sounds flat spectrums. ^ 4 cl, 4 dwg |
