JPH0314360B2 - - Google Patents

Description

[Detailed Description of the Invention] This invention provides speaker verification that determines whether or not the input speaker is the registered speaker by determining the degree of similarity between the input voice and the stored voice. It is related to the device.

FIG. 1 shows an example of the configuration of a conventional speaker verification device. The operation will be explained below using figures. It is assumed that the reference pattern storage circuit 1 stores word sounds uttered by registered speakers in advance in the form of feature parameters after being analyzed in the same manner as the feature parameter extraction circuit described later. The speaker inputs a registration number into a registration number input circuit 2, such as a magnetic card reading circuit or a keyboard input circuit, and calls up his or her registered reference pattern 3. Next, face the microphone 4 and say the determined word. The audio signal 5 is subjected to, for example, spectrum analysis in a feature parameter extraction circuit 6 and converted into feature parameters such as cepstrum at regular time intervals, and an input pattern 7 is extracted. Since the input pattern 7 and the reference pattern 3 have different durations, a DP matching circuit 8 performs matching using nonlinear time expansion and contraction to find the distance 9 between the two patterns. The distance 9 is compared with a threshold value in a determination circuit 10, and a recognition result 11 is output in which if it is less than the threshold value, it is accepted as the real person, and if it is more than the threshold value, it is rejected as an impostor. However, since DP matching minimizes the distance between two patterns by nonlinear expansion and contraction, it is not suitable as a matching method for a speaker matching device that aims to find differences between speakers. In other words, if the time structure within a word that is unique to a speaker is expanded or contracted more than necessary, differences between speakers often cannot be reflected sufficiently in the distance, and if a limit is set on time expansion or contraction to avoid this, it is necessary to In severe cases, there is a risk that matching between utterances of the same speaker may result in correspondence between allophones. Furthermore, when making a determination using the average distance of all word sections obtained by DP matching, there is a possibility that differences in a specific phoneme within a word may be overlooked because other phonemes are similar. Conventional speaker verification devices have these drawbacks, which are the cause of a decline in recognition rate.

This invention was made to improve these shortcomings, and it is possible to sufficiently reflect differences in temporal structure within a word that are unique to speakers, without overlooking specific phonological differences within a word. The purpose is to provide a device that is capable of making judgments and requires a small amount of calculation.

FIG. 2 shows an example of the structure of this invention. In the figure, 12 is a segmentation circuit that cuts out a predetermined time interval from an input pattern, 13 is an interval input pattern, and 14
15 is a section reference pattern, 15 is a section reference pattern storage circuit, 16 is a linear matching circuit, 17 is a section distance, 18 is a section judgment result, and 19 is a logic judgment circuit.

Next, the operation will be explained. It is similar to conventional devices until the speaker inputs his or her registration number, utters the words, and is converted into feature parameters. The characteristic parameters of the input speech are subjected to detection of phoneme change points and the like in a segmentation circuit 11, and are segmented into predetermined interval input patterns 12. Since the words to be input are predetermined, segmentation can be performed relatively easily and with high precision using the amount of change in feature parameters over time.

FIG. 3 shows an example of a procedure of processing in the partitioning circuit 12. First, in A, the distance D(m) between adjacent frames of the input pattern 7 (m is the frame number), that is, the amount of change over time of the parameter, is determined, and then in B, the peak of D(m) is detected.

The peak value Dp is compared with the threshold value Dt at
If it is above the threshold, the frame number is stored as a phoneme change point in E. If it is less than the threshold, the process returns to (a) and detects the phoneme change point repeatedly until the end point of the word is detected in (c). The section input pattern 13 is cut out from the input pattern 7 using the phoneme change points determined above as the starting point and the ending point, and outputs it. On the other hand, the section reference pattern storage circuit 15 stores section reference patterns 14 that have been segmented in advance, and is sent to the linear matching circuit 16 by a signal from the registration number input circuit 2. A linear matching circuit 16 calculates a section distance 17 by normalizing the duration of the section input pattern 13 and the section reference pattern 14 by linear expansion and contraction.

In other words, the interval input pattern is {t _o } (n=0,
1,...,N-1), the interval reference pattern is {r _n }
(m = 0, 1, ..., N-1), where n and m are frame numbers. Distance D by linear matching
is calculated as follows. Pattern {t _n } (m=0, 1, ...M-
Find 1).

t _n = a×tt _k + (1-a)×tt _k+1 k=[N-1/M-1×m] a=N-1/M-1×m-k Here, [i] is A Gaussian symbol indicates the integer part of i.

The distance D is D=1/M _M-1 〓 ⁱ⁼⁰ {dr _i , t _i } Here, d(r _i , t _i ) is the i-th frame of {1r _n } and the i-th frame of {tt′ _n } Indicates the distance from the eye. The determination circuit 10 performs threshold value determination for these section distances 17. Furthermore, these section determination results 18 are logically determined by a logic determination circuit 19 to obtain a final recognition result 11. For example, section judgment result 18 is A ₁ ,
A ₂ ...A _l ...A _L (L: number of sections), A _l = 1 when the section distance is less than the threshold, A _l when it is greater than the threshold
= 0, and if the recognition result S is obtained by logical product, then S = A ₁・A ₂・……・A _l・……A _L , that is, only when the interval distance is within the threshold in all intervals It will be accepted as the real person, and if it exceeds the threshold even in one section, it will be rejected as an impostor.

As explained above, according to the present invention, it is possible to sufficiently reflect the temporal structure in the distance by linear matching, and if the interval is within about three phonemes, it is possible to Almost no correspondence occurs. Furthermore, since the recognition results are obtained by logically determining the threshold value results for each section, there is less possibility of overlooking specific phoneme differences within a word, and a higher recognition rate than before can be obtained. Furthermore, the amount of calculation is
Since DP matching is not performed, it is considerably smaller, and compared to the conventional method, the added segmentation circuit also requires a relatively small amount of calculation because the words are determined.

Note that in the above embodiment, the linear matching circuit 16,
Although the determination circuit 17 is configured in parallel for each section, it is also possible to sequentially output the section patterns from the sectioning circuit 12 and the section reference pattern storage circuit 15 and process them with one linear matching circuit and one judgment circuit. Needless to say.

[Brief explanation of the drawing]

FIG. 1 is a diagram showing a conventional speaker verification device, FIG. 2 is a block diagram showing an embodiment of the speaker verification device according to the present invention, and FIG. 3 is a diagram showing an example of segmentation in the segmentation circuit in FIG. 1 is a diagram showing a procedure example, 1 is a reference pattern storage circuit, 2 is a registration number input circuit, 3 is a reference pattern, 4 is a microphone, 5 is an input voice, 6 is a feature parameter extraction circuit, 7 is an input pattern, and 8 is a diagram showing an example of a procedure. is the DP matching circuit, 9 is the distance, 10
is a judgment circuit, 11 is a recognition result, 12 is a segmentation circuit, 13 is a section input pattern, 14 is a section reference pattern, 15 is a section reference pattern storage device, 16 is a linear matching circuit, 17 is a section distance, and 18 is a section judgment As a result, 18 is a logic judgment circuit. It should be noted that the same or corresponding parts in the figures are indicated by the same reference numerals.

Claims (1)

[Claims] 1. A feature parameter extraction circuit that spectrally analyzes the input word speech of the target speaker, converts it into feature parameters at regular time intervals, and extracts the input pattern, and the feature parameters extracted by this feature parameter extraction circuit. A segmentation circuit detects phonological change points within a word duration from an input pattern, divides it into multiple time segments, and extracts the characteristic parameters of each segment, and segments the word sounds uttered by registered speakers in advance. an interval reference pattern storage circuit which stores the interval reference pattern in a state that can be called up by specifying the registration number from the registration number input circuit; a linear matching circuit for correlating the pattern with each of the plurality of section reference patterns specified by the registration number input circuit and called from the section reference pattern storage circuit by normalizing the duration by linear expansion and contraction, and calculating the distance of each of the plurality of sections; A judgment circuit that judges pass/fail for each plurality of sections with respect to a predetermined threshold value for each of the plurality of section distances obtained by the linear matching circuit, and a logical judgment is made by integrating the plurality of section threshold judgment results of this judgment circuit. and a logical determination circuit for determining whether or not the speaker of the input word voice is the person registered in advance.