SU1675936A1 - Method for verification of speaker - Google Patents

Abstract

The invention relates to the field of automatic pattern recognition based on the analysis of speech signals and can be used to automate checkpoints and develop voice keys that allow authorized persons to access data banks. The method simplifies the verification process and improves its speed while maintaining reliability. The method consists in pronouncing by the speaker a control phrase containing a sequence of deaf and single-phone voiced sections, highlighting voiced sections and filtering them in the frequency bands of the main voice of the voice and the third and fourth formant, determining the sequence of values of the densities of zero signal intersections in the filtered frequency bands and threshold comparison of the parameters of the selected features with reference independently within each selected area from its beginning. 2 Il.

Description

The invention relates to automatic pattern recognition based on the analysis of speech signals and can be used to automate checkpoints of agencies and develop voice keys that allow authorized persons to access data banks or special purpose mechanisms based on the presentation of personality of his speech.

The aim of the invention is to simplify the verification process and increase its speed while maintaining a high level of verification reliability.

FIG. 1, using a specific example of the control phase (words), gives a timing chart for the analysis of acoustic material; FIG. 2 shows a block diagram of a device implementing the method.

The method is illustrated by a specific example of the analysis of the control phrase (word) of the poster (Fig. 1).

FIG. 1 and shows the oscillogram of the speech signal; in fig. 16 is a speech signal that is segmented by voicing-deafness.

Within the selected voiced sections, the beginning and end of which determine the control pulses (Fig. 1c), at the times corresponding to the clock pulses (Fig. 1d), the values of the densities of zero intersections of the signals corresponding to the highlighted parts of the speech segment in filtered frequency bands of the main tone of the PQ voice (Fig. 1e). third pi and fourth pi

ABOUT

XJ ave you

Os

formant (fig. 1e and g, respectively). These signs characterize the speaker's individual manner of speaking, as well as the tonal and tonal features of his voice. According to the discrete values of the signs, ro (+) i / 9i (+), (+) form the data matrix (Fig. 1h). The data matrix contains submatrices that are compared with the corresponding reference submatrices separately, the number of rows of all three submatrices (in the specified example) is three, the number of columns in the first submatrix is four, in the second six and in the third four.

The resulting matrix is to be compared with the reference one in accordance with the pre-selected value of the comparison threshold 9.

The device implementing the method (Fig. 2) consists of an amplifier 1, a speech signal splitter 2 into voiced and deaf sections, low-pass filters 3 for separating the first signal harmonic and two band-pass filters 4 and 5 separating the spectral regions corresponding to the third and fourth formants. Blocks 6-8 carry out the selection of the sign of the density of zero intersections p0, pt, speech signal passing, respectively, through the filters 3-5. Data on changes in densities / e zeros in time are recorded in a storage device — shaper 9 of a data matrix as discrete values at discrete points in time. The recording of the corresponding data in the block 9 is controlled by the block 10, which fixes the beginning and the end of the ringing sections of the signal, and the block 11, which generates in these time intervals the clock pulses sampling values.

The device operates in the learning mode - the right position of the switch 12, and verification - the left position of the switch 12,

The data from block 9 in the learning mode is transmitted to block 13, which is a memory device. In the verification mode, data on unknown pronunciation is recorded in block 14, which is a comparing device that compares data from blocks 13 and 9. Decision unit 15 analyzes the results of the comparison and states the fact or non-ownership of the control pronunciation of the reference.

The device works as follows.

The speech signal, after amplification to the magnitude necessary to support the operation of the following blocks, is fed to separator 2, which contains a set of

narrowband filters covering the frequency range of 100-500 Hz. The bandwidth of each filter is not more than the possible pitch frequency of the voice and is approximately 70-80 Hz. In time

the device makes it possible to determine the line structure of the signal and separate it from the continuous spectrum that occurs when there is a deaf sound.

Selected blocks 2 voiced sections

speech is given to three filters. The first of these (3) is a low-pass filter with a decay slope of 18 dB / oct, starting at 200 Hz. This filter serves to highlight the first harmonic of the signal. Filter 4

is a band-pass filter in the spectral region of the third formant at 2200–3200 Hz. Filter 5 is also a bandpass filter in the fourth formant area of 2800–4000 Hz.

The signals passing through these filters are then fed to blocks 6, 7 and 8 (/ o-meters), which form a sequence of values of the zero points of the signal, p-meters are

simple capacitor drives or integrators that produce certain potentials depending on the number of zero points of the signal arriving at them in the interval between two clock pulses.

Obviously, the higher the frequency of the speech signal passing through the filters 3, 4 and 5, the greater the number of zero points and, therefore, the higher the potential at the drive in

within two adjacent clock pulses. The measured values at the outputs of blocks 6-8 reflect both the individual characteristics of the fundamental tone of the voice and the two higher formant, as well as the peculiarity of their change in time at discrete points in time, due to the clock pulses. The sequences of these discrete values reflect at the output of block 6 the character of the change (melody) of the pitch of the voice, at the outputs of blocks 7 and 8 the character of the change of the third and fourth formant of speech.

To form a data matrix, the discrete values of all the selected

signs (pa, pi, and / for) are recorded in a storage device (block 9) with strictly defined addressing. Controlling the process of sampling the values y ° o (t) ./9l(t) and / ° 2 (g) P and recording these values

in the memory unit, block 10 records the beginning and end of voiced signal sections and, thus, determines the write address of the corresponding data in 3, and block 11, which is a clock pulse generator, used to convert the analog characteristic values to 25 Hz at a frequency of 25 Hz in the interval between the control pulses (block 10), the data matrix thus formed serves as the basis for characterizing the human voice. In the learning mode (right position of the switch 12) the data matrix is transferred to the non-volatile memory (block 13) in the verification mode (left position of the switch 12) this data is transmitted to block 14, where they are compared with these blocks 13. Block 14 is actually an arithmetic unit that calculates the difference between two matrices. In this case, the difference is taken separately for different parts of the matrix, corresponding to submatrices. Each submatrix reflects the corresponding voiced portion of the spoken speech segment. If the number of submatrices is unequal in the control pronunciation and the standard, a negative signal is output to the decision block 15 indicating a foreign pronunciation. When the number of submatrices is equal, the decision block analyzes the value of the matrix difference and, depending on whether the absolute value of the difference between the matrices of a certain threshold 9 obtained in the comparison block, indicates that the speaker is native or foreign, respectively. As threshold 0, 10-20% of the sum of all members of the reference matrix is taken. In the case of a decision, its block 15 generates an output impulse at the output.

In case of inequality of the number of columns of the control submatrix and the reference

submatrices comparison is performed for those columns that have nonzero values in both submatrices, i.e., the comparison occurs by the shortest (by the number of columns) matrix (by significant

Claims (1)

non-zero matrix columns). The invention of the method of verification of the announcer, which consists in the speaker pronouncing a pre-selected control phrase, identifying the acoustic features of a speech signal, determining their parameters and threshold comparison of the parameters of the selected features with predetermined reference values, from which it is necessary to simplify and increase speed by simplifying the correlated signals, ensuring minimization of errors without use operations of nonlinear normalization of signals, use a speech segment as a control phrase, with a sequence of deaf and single-telephonic voiced sections, the selected speech segment is segmented into voiced and deaf sections, the sequence of values of zero-intersection signals of the signals corresponding to the selected voiced sections of the speech segment in the filtered frequency bands is determined pitch of the voice and the third and fourth formant, and the threshold comparison of the parameters of the selected features is made by hundred tion within each selected area. aPtvtlou signal 1675936 , ftCememu.uptit - lima and gmein cosiness ptvetoa and nzla bc / ro & rs intsiki uTahmJ / e upei 5 "L to i | yf; / I g) A fiifu- felMiiftbW hifftt fit fa) J "n" "JW l - fiiflt fiifiifrdttttti Jlln Jnn 1 /four  H fj