JPH0632007B2 - Speaker verification method - Google Patents

Description

DETAILED DESCRIPTION OF THE INVENTION Object of the Invention (Industrial field of application) The present invention reliably verifies a speaker by comparing a pre-registered voice feature pattern with an input voice feature pattern. Speaker verification method for doing.

(Prior Art) Research on speaker verification has been performed for a long time, and various methods have been proposed. As a typical method, BPF (Band Pas
There is a method in which the similarity between voice patterns by s Filter) analysis is obtained by processing such as DP matching, and determination is performed based on this similarity. In BPF analysis, speech is passed through multiple BPFs with different pass frequency bands in parallel, each output is smoothed by LPF after full-wave rectification, and the pattern is obtained by A / D conversion at certain time intervals. DP matching is an efficient method for non-linearly matching pattern lengths so that the distance between two patterns is as small as possible. For BPF analysis and DP matching,
For example, it is explained in detail in "Voice Recognition" published by Kyoritsu Publishing by Yasunaga Niimi.

(Problems to be Solved by the Invention) In the above-mentioned method, since the BPF analysis pattern itself is used as a feature, there is a drawback that the amount of data becomes very large. For example, when analyzing 1 second long speech with 15 BPFs and A / D period of 10ms, the data amount becomes 1000/10 × 15 = 1500.

The present invention has been made under the circumstances as described above,
An object of the present invention is to reduce the total data amount by using a feature amount selected from the feature amount related to the distance pattern with five vowels in consideration of effectiveness, and to reliably match the speaker even with a small data amount. It is to provide a speaker verification method for doing so.

Configuration of the Invention (Means for Solving the Problem) The present invention relates to a speaker verification system, and an object of the present invention is to make a speaker utter five vowels in advance to temporarily create a five vowel pattern, and A word other than the 5 vowels is uttered to obtain the distance from the 5 vowel patterns, an effective one is selected from the feature quantities, and the selected feature quantity is registered as a standard pattern of the speaker. This is achieved by uttering the same words other than the five vowels as at the time of registration at the time of matching, and matching the speaker based on the degree of similarity with the registered standard pattern.

(Function) The present invention was developed to obtain a high speaker verification ability with less data. By selecting and using a feature amount with more individuality, it is 1/10 or less compared to the conventional method. It is possible to collate with the amount of data. In the present invention, in particular, a specific speaker is made to speak 5 vowels and words other than 5 vowels in advance to calculate a feature amount, the feature amount is selected in consideration of effectiveness, and the feature amount is registered as a standard pattern. Every time, the words other than the 5 vowels are uttered at the time of matching to calculate the characteristic amount, and the pattern obtained from the characteristic amount is compared with the standard pattern to perform the speaker verification.

(Embodiment) FIG. 1 shows an operation example of the present invention. First, it is judged whether the mode is the registration mode or the collation mode.
First, the registration mode will be described.

First, a parameter m for counting the number of times of inputting 5 vowel sounds
Is set to "0" (step S2), and then the speaker utters five vowels "a" to "o" (step S3), and acoustic analysis is performed for each vowel of the uttered five vowels. (Step S4). This acoustic analysis is for LPC (Linear Planning Coefifcient) analysis of speech data obtained by A / D converting a speech signal. The details of LPC analysis are explained in "Speech Synthesis and Recognition" by Kazuo Nakata published by Sogo Electronics Co., Ltd. This method is used.

That is, as shown in FIG. 2, a voice waveform is A / D converted at an appropriate sampling time, a voice section is cut out, divided into frame units, for example, with a frame shift of 10 ms and a frame length of 30 ms. The Hamming window is multiplied to eliminate the distortion of the frequency at both ends of the frame, and then the adaptive first-order multistage inverse filter (calculated)
Then, the LPC analysis is performed after removing the one due to the sound source characteristic and leaving only the vocal tract characteristic.

The data pattern is as shown in FIG. Here, a _ij indicates the i-th frame and the j-th LPC coefficient, and n indicates the analysis order. FIG. 4 (A) shows the characteristic after the Hamming window is crossed, and FIG. 4 (B) shows the characteristic after the inverse filter is passed through the same figure (A). The spectrum is flattened by the inverse filter. Further, the LPC time series in the steady state (a portion excluding the beginning and end portions) is averaged as shown in FIG. 3 to first create a vowel pattern "a".
Similarly, LPC analysis is performed for “i” to “o” as well.
Create a vowel pattern.

After this acoustic analysis, it is determined whether or not the parameter m is a predetermined number M (step S5), the parameter m is counted up until m = M (step S6), and the process returns to step S3. Then, the above operation is repeated. When m = M, the patterns M times are averaged,
5 vowel patterns are stored in the memory (step S7),
The parameter l for counting the number of voice inputs is set to "0" (step S8), and a predetermined sentence, word, etc. is uttered (step S10. Note that the average pitch and the average inverse filter coefficient described later are five vowel patterns. Do not store in memory.

For this uttered voice, first, the same acoustic analysis as described above is performed (step S11, the inverse filter coefficient is obtained, the time sequence of the LPC analysis is obtained after passing through the inverse filter, and the time sequence of the pitch is further determined. Seeking Pitch (voice pitch)
It is easier to find the difference between what was predicted by the LPC and the actual data, and then to find the time series of the pitch from this residual, rather than to find it directly from the original data itself. And
A feature amount is calculated based on the acoustic analysis data (step S12). This feature is calculated with the average pitch (pitch of voice), average LPC coefficient (average characteristic of vocal tract), average inverse filter coefficient (average characteristic of sound source) and 5 vowels in LPC pattern. This is to obtain various feature quantities related to distance. The relationship between the LPC coefficient, the pitch and the inverse filter coefficient for each frame and the average LPC coefficient, the average pitch and the average inverse filter coefficient is shown in Fig. 5, and the distance between the five vowels in the LPC pattern is also shown. Becomes matrix data as shown in FIG. Distance to vowel "i" is block SC
Although represented by 1, the distance pattern d _aj with the vowel “a” is calculated by the following equation.

α _jk : L-th coefficient of the j-th frame and the k-th order α ▲ ^a _k ▼: L-th coefficient of the k-th order of the vowel “a” d _aj : Pattern distance from the j-th frame vowel “a” Also, it relates to the distance from the fifth vowel. Feature content is 7th
The calculation formulas for the vowel “a” in the upper row (for each vowel) and the vowels “a” and “i” in the lower row (between each vowel) shown in FIG. Is shown below. The variance value bunt of the entire voice section for each vowel is d _ai : Distance from “a” in i-th frame Average n: number of frames. Also, the correlation coefficient r of the entire voice section between each vowel
Is : Covariance of d _ai and d _ii d _ai of the variance s _ii: is the variance of the _{d ii.} And the distance sum dist is Is. Furthermore, for the j-th segment, the variance value bunsj for each vowel, the average value avesj, and the variance value dis-bu for the distance between vowels.
The average value of nj and distance dis-avej is calculated as follows.

(When the j-th segment includes the m-th to m-th frames) _{co-disj = (d ai -d} ii) 2: Distance d _ai and d _ai FIG. 8 shows the distance from the 5 vowel pattern in the LPC pattern when "namae" is uttered. Characteristic DS1
Is the distance from vowel "a", characteristic DSD is the distance from vowel "i", characteristic DS3 is the distance from vowel "u", characteristic DS4 is the distance from vowel "e", and characteristic DS5 is the vowel It shows the distance with "o", and when the distance between the vowel and the corresponding vowel is small in the vowel part and the five are disjointed, the utterance is clear, and when the five have relatively large and similar values, It can be said that the utterance is unclear.

The relationship between the features of FIG. 7 is shown in FIG. 9 (A),
It is (B), and (A) of the same figure extracts two characteristics from FIG. The vertical solid lines in FIGS. 9A and 9B indicate the boundaries of the segments. The distance from the vowel “a” to the frame (time) is shown by the characteristic A in FIG. 9 (A), and the distance from the vowel “i” is shown by the characteristic B.
The distance sum (dist) between the vowels “a” and “i” is indicated by the shaded area C. Since the undulation of the distance B is steeper than that of the distance A, the variance value (bunt) in the entire voice section becomes large. The X mark in the figure indicates the minimum value (m
int),-mark indicates the minimum value (mins) in the segment, and the horizontal line indicates the average value (aves) in the segment. The distance between the vowel “a” and the distance between the vowel “i” is expressed as shown in FIG. 9 (B), and the average value (dis-ave) in each segment is indicated by the horizontal line. The minimum value (dis-min) is marked with *.

However, the segment is defined as the vowel stationary part and the space between the vowel stationary part and the beginning and end of the vowel, and the distance pattern with the five vowels is used, and the segment is performed top down in consideration of the utterance content. For example, when "namae" is uttered, in order to find the first "a" segment, the minimum value of the distance from "a" in the existence candidate section (before 1/3 of the total length in the case of "namae") Find the threshold value, make it the minimum value x a constant, and make the length of the section equal to or less than the threshold value within the range of the segment length (predetermined range depending on the uttered words, for example, 10 to 20 frames). Calculate the segment by adjusting the constant value to.

After such calculation of the feature amount, the process returns to step S10 and the same operation is repeated until the parameter l reaches a predetermined number L (steps S13 and S14). Selection is made (step S15). That is, ten valid voices are selected from FIG. 7 for each speaker and voice. The effectiveness is evaluated by the magnitude of the F ratio shown by the following formula.

Where μ _i : center of distribution of speaker i μ: center of the whole x _ij : j-th data of speaker i I: number of speakers L: number of data The smaller the spread of the principal distribution, the farther the distribution of each speaker is, the better the separation state is, and the larger the inter-speaker distribution is, The smaller the dispersion, the larger the F ratio and the more effective. In addition, the average pitch, average LPC coefficient (3rd to 8th order), and average inverse filter coefficient (1st to 3rd order) are commonly used regardless of the speaker's vocal sound.
Adopt 10 pieces. That is, the number of feature quantities for matching is 20 in this embodiment. In addition, 5 as shown in FIG.
It has been confirmed by experiments that it is very effective for speaker verification if at least 10 of the various features related to the distance from the vowel pattern have a large F ratio, so 10 are selected. The number of selections may be increased slightly depending on the calculation processing time.

After such feature amount selection, L feature averages of the selected feature amounts are averaged and registered in the memory as a standard pattern (step S16). However, since the above-mentioned calculation of the F ratio requires data other than that of the person himself / herself, it is necessary to collect some data other than that of the person who should be compared before the registration.

After the registration mode as described above, the speaker verification is performed as follows. That is, first, the predetermined sentence used at the time of registration is uttered in words (step S20), and the input voice is subjected to the acoustic analysis as described above (step S20).
S21) and calculation of the feature amount selected at the time of registration (step S2
2) is performed, and then the average pitch is determined (step S23). The determination based on the average pitch is based on whether the average pitch in the standard pattern is within a certain range. That is, the determination in _{f oref × (1-α)} ≦ f oin ≦ f oref × (1 + α) ...... (10) f oref: Mean standard pattern pitch f _oin: when the average pitch of the input voice is OK and Become. After the determination based on the average pitch, the determination based on the weighted distance is performed (step S24). The determination based on the weighted distance is performed by comparing the weighted distance according to the effectiveness of 19 types of feature values other than the average pitch with the threshold value according to the following formula.

w _i : weight of the i-th feature (= F ratio) a _i : value of the i-th feature of the standard pattern b _i : value of the i-th feature of the input pattern d _is ≦ d _iso OK d _is > d _iso NG As described above If the determination based on the average pitch is OK and the determination based on the weighted distance is OK, the speaker is accepted as being the same as the registrant (step S25), and if either is NG, it is rejected. (Step S26).

EFFECTS OF THE INVENTION As described above, according to the speaker verification method of the present invention, since effective ones are selected and used from the feature quantities related to the distance pattern with the five vowel patterns, a small amount of data and a speaker can be used. It is possible to reliably perform the collation of.

[Brief description of drawings]

FIG. 1 is a flow chart showing an operation example of the present invention, and FIG.
Figures and 3 are figures for explaining LPC analysis, and Figure 4
(A), (B) is a diagram for explaining the relationship of the inverse filter, the fifth
The figure shows the contents of the feature amount, FIGS. 6 and 8 show the distance pattern with 5 vowels, FIG. 7 shows the contents of the feature amount, and FIGS. 9 (A) and 9 (B). FIG. 4 is a diagram for explaining a feature amount.

Claims (1)

[Claims] 1. A speaker utters 5 vowels in advance to temporarily create a 5 vowel pattern, and a word other than the 5 vowels is uttered to obtain a distance from the 5 vowel pattern, which is effective from the feature quantities. , The selected feature quantity is registered as a standard pattern of the speaker, and at the time of verification, the same words other than the 5 vowels as at the time of registration are uttered, and the registered standard pattern A speaker verification method characterized by matching speakers based on the degree of similarity.