JPS6131479B2 - - Google Patents

Description

DETAILED DESCRIPTION OF THE INVENTION The present invention relates to a speaker verification method, and more particularly to a method for automatically determining with high precision and efficiency whether or not a particular speaker among registered speakers is the same person.

As one of the telephone service systems, there is a system that determines whether or not the voice of a particular person among the registered speakers is actually that of that person, and notifies the user of the result. This allows it to be used for shopping over the phone, depositing and withdrawing money, identifying the culprit who made the threatening call, or conducting transactions, meetings, etc. over the phone.

Speech waves are sound waves emitted from the lips or nose when vocal cord vibration waves (voiced sound source) and noise waves due to turbulence generated by the narrowing of the vocal tract (unvoiced sound source) are applied to the vocal tract.

FIG. 1 is a block diagram electrically modeling a sound wave.

The voiced sound from the impulse generator 21, which is the vocal cords, and the unvoiced sound from the white noise generator 22, which is the turbulent flow, are
It is switched by a switch 23 and connected to an electric circuit 24 representing a vocal tract pipe, causing a speaker 25 to produce a sound wave. In FIG. 1, the left side of the chain line is a sound source, and the right side is a vocal tract characteristic.

When analyzing audio, there is a certain finite amount (10 to 30 mS)
The objective is to quantitatively clarify the frequency spectrum of the speech wave (the spectrum of the transfer function of the vocal tract) at each moment of the sound wave and the properties of the sound source that drives the vocal tract. It's called analysis.

A sound source is expressed by three elements: a classification signal (presence/absence) V indicating whether it is impulse train driven or noise driven, the period (pitch) P if it is an impulse train, and the amplitude A of the impulse train or noise. It can be done. Sound source analysis involves quantitatively extracting these three elements, but since these elements change at a fairly high speed, accurate analysis is quite difficult.

FIG. 2 is a block diagram showing a speech analysis model.

In order to analyze the sound source of a voice, the voice is input to the Percoll analyzer 26, a spectrum analysis is performed in advance (K ₁ , K ₂ . . . K _p ), and the voice wave is passed through an inverse spectrum circuit having inverse characteristics of the spectrum. 28
to eliminate peaks and valleys in the spectrum. At this time, the obtained spectrum becomes substantially flat, and the impulse train or noise of the sound source appears as a residual wave R, and the residual wave R contains a sound wave signal.

Therefore, the sound source analysis circuit 29 analyzes the residual wave R and extracts the sound source signals V, P, and A.

By the way, the device used in the speaker verification method determines in advance the utterance content of the input voice used for determination, that is, the words, and then always utters the same words and compares them with the registered voice waves of each speaker. One method uses arbitrarily uttered words to determine whether or not the input speech can be considered to be the voice of the speaker, and the other method uses arbitrarily uttered words to identify only the characteristic parts of the input speech without predetermining the utterance content of the input speech used for determination. It can be divided into two types: one that is determined by comparing it with the characteristic parts of the voice waves of each registered speaker.

The latter method generally has a wider range of applicability, but it is difficult to make a highly accurate determination. On the other hand, although the former method has somewhat limited uses, it is expected to have an extremely wide range of practical applications, and it is possible to obtain higher accuracy than the latter method.

Generally, it is not efficient to directly compare input speech waves and registered speech waves, so it is desirable to convert them into so-called feature parameters such as frequency spectra and linear prediction coefficients before comparing them. In addition to the above, the conventional configuration of this type of device includes fundamental frequency, voice energy, formant frequency, cepstral coefficient, Percoll coefficient, logarithmic cross-sectional area ratio,
The number of zero crossings, etc. is used, but it is difficult to extract parameters stably and accurately, complex calculations are required to extract the parameters, and parameters that do not express the characteristics of the speaker's voice are inadequate. However, there are drawbacks such as the fact that the accuracy of the determination is greatly reduced due to fluctuations when passing through a transmission line such as a telephone system.

The purpose of the present invention, in order to eliminate such drawbacks, is to easily extract voice characteristics that are not easily affected by transmission distortion etc. from voice transmitted through a telephone system, etc., and to determine with high accuracy whether or not the person is the real person. The object of the present invention is to provide a speaker verification method that can perform the following steps.

The speaker verification method of the present invention uses linear predictive cepstral coefficients, which are extremely useful and can be extracted by a relatively simple method, as parameters expressing the characteristics of the speaker's voice. Extracts polynomial expansion coefficients, which are characteristic parameters that are not easily affected by fluctuations, and determines whether or not the voice is by the person himself/herself through nonlinear time normalized matching (correlation) with the characteristic parameters of each speaker registered in advance. It is characterized by

Hereinafter, an embodiment of the present invention will be described with reference to FIG.

FIG. 3 is a block diagram of the speaker verification method of the present invention.

As shown in FIG. 3, the method of the present invention involves inputting the speech to be verified from the speech input terminal 1, and then inputting the speech to be verified using the speech section detection circuit 3, linear prediction analysis circuit 4, cepstrum conversion circuit 5, cepstrum storage section 6, Polynomial expansion coefficients are extracted from the time waveform of the linearly predicted cepstrum coefficients via a cepstrum averaging circuit 7, a subtraction circuit 8, a feature parameter storage section 9, a cepstrum register 10, and a polynomial expansion circuit 11.

On the other hand, the number of the speaker to be verified is inputted from the identification number input terminal 2, the corresponding pattern is retrieved from the standard pattern storage section 13, the learning mode and the matching mode are switched by the switch 12, and the voice to be verified is inputted from the identification number input terminal 2. The nonlinear time normalization circuit 15 calculates the degree of similarity by outputting the polynomial expansion coefficient of The standard pattern averaging circuit 1 outputs an output indicating whether the
4, and the threshold value used to determine the speaker is input to the threshold calculation logic circuit 19 and updated.

The operation will be explained in more detail. First, an audio wave used for speaker determination is inputted from the audio input terminal 1, and at the same time, the number of the speaker to be verified is inputted from the identification number input terminal 2. To input this number, for example, a dial on a push-button/dial telephone can be used. Since the input speech wave usually includes an actual speech section and a silent (noise) section, the input speech wave is input to the speech section detection circuit 3 to detect the speech section. .

For this detection, several well-known methods can be used, such as the short-term energy of the input signal wave, the time that energy above a certain value continues, the presence or absence of periodicity of the waveform, etc. The signal wave of the detected voice section is sent to the linear prediction analysis circuit 4 and converted into a time waveform of a linear prediction coefficient. This technique is already known (for example, see the literature, Itakura, Saito: Estimation of speech spectral density and formant frequency by statistical methods, Journal of the Institute of Electronics and Communication Engineers, 53-A, 1, 35, 1970).
Although the details are omitted, basically the waveform is first passed through a low-pass filter, then sampled and digitized, a short section of the waveform is cut out at regular intervals, multiplied by a Hamming window, etc., and then calculated using a product-sum calculation. and calculate the correlation coefficient. For example, the length of the humming window is 30 m.
S, the period for updating this is, for example, 10m.
A value such as S is used. Linear prediction coefficients are easily extracted from the correlation coefficients by solving algebraic equations through iterative calculation processing. The correlation coefficient and linear prediction coefficient are, for example, from the 0th order to the 10th order.
Calculate values up to. The time waveform of the extracted linear prediction coefficient is converted by the cepstrum conversion circuit 5 to
It is converted into so-called linear predictive cepstral coefficients. The linear predicted cepstral coefficients are the logarithmic power
Although it is slightly different from the conventional cepstral coefficients obtained by Fourier transform of the spectrum, the spectral envelope expressed thereby is similar.
It is known that the linear predictive cepstral coefficient has excellent properties as a parameter expressing the voice characteristics of the speaker (Reference, BSAtal:
Effectiveness of Linear Prediction
Characteristics of the Speech Wave for
Automatic Speaker Identification and
Verification, J.Acoust.Soc.Amer., 55, 6,
p. 1304, 1974). Conversion from linear prediction coefficients to linear prediction cepstrum can be performed by the following calculation.

_C1 = _a1 ...(1) Here, a _o is an n-th linear prediction coefficient, c _o is an n-th linear prediction cepstrum, and p is the number of dimensions of the linear prediction model. As described above, a value of about 10 is used as p.

The time waveform of the linearly predicted cepstrum coefficients (hereinafter simply referred to as cepstrum coefficients for simplicity) of the extracted entire speech section is stored in the cepstrum storage unit 6. At the same time, the waveform of the coefficient, which is predetermined as a characteristic parameter to be used later for speaker determination, is input to the cepstrum averaging circuit 7. Here, the average value of the entire speech interval is calculated for each cepstral coefficient of each degree.
Which coefficients among all cepstral coefficients are used as feature parameters is determined in advance through preliminary experiments and statistical analysis such as variance analysis.

Next, this average value and the waveform of the coefficient determined to be used as a characteristic parameter among the time waveforms of the cepstrum coefficients stored in the cepstrum storage unit 6 are input to the subtraction circuit 8, and the value of each cepstrum coefficient is Subtract the average value of the corresponding order from . This output is temporarily stored in the feature parameter storage section 9.

On the other hand, among the cepstrum coefficients stored in the cepstrum storage unit 6, the time waveforms of a plurality of predetermined coefficients are temporarily stored in the cepstrum register 10 in sections of a certain time length at certain intervals, respectively. The contents of this register 10 are sent to a polynomial expansion circuit 11 to calculate polynomial expansion coefficients. This cepstrum register 1
0 and the length of the time waveform of the cepstrum coefficients input to the polynomial expansion circuit 11, for example,
90mS, and the update period is, for example,
Use a value like 10mS. Various methods can be used to expand the time waveform into a polynomial. Here, for example, a method is used in which the time waveform is expressed as a linear combination of the following three types of functions.

P _0j = 1 ... (3) P _1j = j-5 ... (4) P _2j = j ² -10j + 55/3 ... (5) At this time, the time waveform of the cepstrum coefficient is x _j
If expressed as (j = 1, 2..., 9), the expansion coefficients corresponding to the above three types of functions are: It can be found by the calculation. Among the coefficients a, b, and c, the coefficients that are predetermined to be used later as feature parameters according to the cepstral coefficients of each order are calculated based on the input of the polynomial expansion circuit 11, which is updated every 10 mS. , and are sent to the feature parameter storage section 9 and stored therein. Among these, the coefficient of a, that is, the 0th-order polynomial expansion coefficient is
Since it corresponds to the short-time average value of the time waveform and is easily affected by fluctuations in the transmission path, etc., it is not stored in the characteristic parameter storage section 9 and will not be used as a characteristic parameter thereafter. The polynomial expansion coefficients b and c express the slope and curvature of the time waveform, respectively.
Since the influence of temporally slow fluctuations in the transmission path, etc. has already been removed as a zero-order expansion coefficient, it has the characteristic that it is not easily affected by the transmission path, etc. The feature parameter storage unit 9 stores time waveforms of a total of about 18 to 20 cepstral coefficients and polynomial expansion coefficients of a predetermined order in the entire speech interval. Among these 18 to 20 time waveforms, the average value of the entire speech interval has already been subtracted from the time waveform of the cepstral coefficient.
Since zero-order coefficients are removed from the polynomial expansion coefficients, both have the characteristic of being less susceptible to the influence of transmission paths and the like. At regular intervals (e.g. 10m as mentioned above)
The 18 to 20 coefficients for each S) are collectively called feature parameters.

The switch 12 is a switch for selecting a learning mode and a matching mode, and for any speaker, for the first utterance, the switch 12 is connected to the terminal 12a, and the feature parameter storage section 9 is input to the standard pattern storage section 13 and stored as the standard pattern for that speaker. For subsequent voices to be determined as to whether the speakers are different or different, first connect the switch 12 to the terminal 12c.
The contents of the feature parameter storage section 9 are input to the nonlinear time normalization circuit 15. At the same time, the standard pattern corresponding to the speaker number input from the identification number input terminal 2 is read from the standard pattern storage section 13 and input to the nonlinear time normalization circuit 15. The nonlinear time normalization circuit 15 calculates the degree of similarity between the standard pattern and the feature parameters of the input voice. The rate of speech production changes both partially and completely each time the same word is uttered repeatedly by the same speaker, so in order to compare the two, it is necessary to , it is necessary to appropriately expand or contract one time axis non-linearly to match the other time axis, and compare the feature parameters at corresponding points in time.

Optimization techniques using dynamic programming can be used as a technique for nonlinearly expanding or contracting the time axis of the other to best fit the two (so that the degree of similarity between the two is greatest) based on one standard. (Reference: Sakoe, Chiba: Continuous word recognition based on temporal normalization of speech using dynamic programming, Journal of the Acoustical Society of Japan, 27, 9, P. 438, 1971).

Also in the method of the present invention, the nonlinear time normalization circuit 15 performs dynamic programming calculations. The feature parameters at a certain point k of the standard pattern are r _ki
(1iN), and the feature parameter at a certain point l of the input audio is expressed by z _li (1iN). Here, the following values are used as the distance between the two (a numerical value indicating that the smaller the degree of similarity is).

or, Here, N is the number of dimensions of the feature parameter, which is a combination of cepstrum coefficients and polynomial expansion coefficients, and uses a value of about 18 to 20 as described above. That is,
Both zli and rki have cepstral coefficients and polynomial expansion coefficients as elements. w _i is a numerical value indicating a predetermined weight for each feature parameter, and this value is based on the result of examining the degree of variability of that parameter using voices uttered multiple times by multiple speakers. Established on the basis of
It is stored in the weight register 16. The distance between the standard pattern and the characteristic parameters of the input voice at corresponding points in time is determined by calculating the distance between the characteristic parameters of the standard pattern and the input voice over the entire voice interval when the time axes are matched so that the degree of matching between the standard pattern and the input voice is the best through dynamic programming calculations. Calculate the average value for. This value will be referred to as the overall distance between the input voice and the standard pattern.

Next, this comprehensive distance and a certain threshold value stored in advance in the standard pattern storage section 13 are inputted to the comparison circuit 17, and the magnitude relationship between the two is determined by a logic circuit. The standard pattern storage unit 13 stores, for each registered speaker, the history of the distance between the standard pattern and the input speaker of that speaker, and the relationship between the standard pattern of that speaker and the input voice of other speakers. A predetermined threshold value is stored based on the distribution of the distance between If the total distance between the input voice and the standard pattern is greater than the threshold, a signal is output from the output terminal 18 to determine that the input voice does not belong to the speaker, and the total distance is determined to be greater than the threshold. If it is smaller than the threshold, a signal is output from the output terminal 18 to determine that the input voice is that of the speaker.

If the input voice is determined to be that of the speaker, connect the switch 12 to the terminal 12b,
The feature parameters stored in the feature parameter storage section 9 are input to the standard pattern averaging circuit 14. At the same time, the characteristic parameters of the standard pattern of the speaker are read from the standard pattern storage section 13,
The correspondence between the standard pattern calculated by the nonlinear time normalization circuit 15 and the time axis of the input audio, that is, the standard pattern average along with a numerical string indicating which time point on one time axis corresponds to which time point on the other time axis. input into the conversion circuit 14. In response to these inputs, the standard pattern averaging circuit 14 calculates a weighted average value of the standard pattern and the input voice for each characteristic parameter at each point in time of the standard pattern. This weight is determined according to the number of input voices of each speaker that have been used so far to create a standard pattern for that speaker. The weighted average value of the feature parameters thus calculated is transferred to the standard pattern storage section 13 and stored as a new standard pattern. Furthermore, the nonlinear time normalization circuit 15
The total distance calculated in step 1 is inputted to the threshold arithmetic logic circuit 19 used for determining the speaker, and the threshold value is updated. As the initial value of the threshold value, a value determined empirically is stored in the standard pattern storage section 13 and used. After that, the threshold value calculation logic circuit 19 receives the sum total of the past two times for each speaker. The target distances are stored, the maximum value including the newly calculated total distance is selected, and the average value of this value plus a constant value and the current threshold value is calculated. This value is
It is transferred to the standard pattern storage section 13 and stored as a new threshold value.

This structure results in speaker verification that shows high accuracy not only for high-quality microphones, but also for voices transmitted through telephone systems, voices affected by noise and transmission distortion, etc. system can be realized. According to previous experiments,
It has been shown that by applying the method according to the present invention to voice transmitted through a telephone set and exchanger including an actual carbon transmitter, speaker verification can be determined with an accuracy of over 99%. There is.

As explained above, according to the present invention, by extracting and using voice characteristics that are not easily affected by transmission distortion etc. from voice transmitted through a telephone system, etc., it is possible to determine whether or not the person is the real person with high accuracy. Therefore, it can be widely applied to various services such as banking services that use telephone voice etc. as a key to determine whether the person is the person who is the user.

[Brief explanation of the drawing]

FIG. 1 is a block diagram showing an electrical model of a speech wave, FIG. 2 is a block diagram showing a speech analysis model, and FIG. 3 is a block diagram showing a speaker verification method according to an embodiment of the present invention. 1: Audio input terminal, 2: Identification number input terminal,
3: Voice section detection circuit, 4: Linear prediction analysis circuit,
5: Cepstrum conversion circuit, 6: Cepstrum accumulation section, 7: Cepstrum averaging circuit, 8: Subtraction circuit, 9: Feature parameter accumulation section, 10: Cepstrum register, 11: Polynomial expansion circuit, 12:
switch, 13: standard pattern storage section, 14: standard pattern averaging circuit, 15: nonlinear time normalization circuit, 16: weight register, 17: comparison circuit, 1
8: Output terminal, 19: Threshold calculation logic circuit.

Claims (1)

[Claims] 1. Means for calculating the time waveform of linear prediction coefficients of the speech wave input to be matched, converting the linear prediction coefficients into cepstral coefficients and storing them, and calculating the average value of the cepstral coefficients in all speech intervals, and calculating the average value and the means for normalizing the cepstrum coefficients by subtracting the time waveform of the cepstrum coefficients, means for calculating polynomial expansion coefficients from the time waveform of the cepstrum coefficients, means for accumulating standard patterns for each registered speaker, and nonlinear time normalization. and a comparison means, the time waveform of the normalized cepstral coefficient;
The polynomial expansion coefficient and the standard pattern extracted from the number of the speaker to be matched are input to the nonlinear time normalization means to calculate the degree of similarity between the two, and the degree of similarity between the two is calculated based on the calculated value and the number of the speaker. A speaker verification method characterized by inputting the extracted value into the comparison means and comparing the magnitude thereof to determine whether the input speech wave is from a speaker corresponding to the number.