JPS61114298A - Speaker collation system - Google Patents

Abstract

(57) [Summary] This bulletin contains application data before electronic filing, so abstract data is not recorded.

Description

DETAILED DESCRIPTION OF THE INVENTION [Technical Field of the Invention] The present invention relates to a speaker verification method used in an information processing system based on human voice input.

[Technical background of the invention and its problems]

In recent years, the development of speech recognition and synthesis technology has been remarkable, for example, continuous speech recognition and speech recognition for unspecified speakers have become possible, and various methods have been devised for speaker verification.

Using such speaker verification technology, applications such as telephone shopping, banking services, access to personal information, and management of people visiting confidential storage locations have been developed, and its usefulness is attracting attention. Here, security is the biggest issue because the problem with these systems is whether or not only one person is using the system. However, current speaker verification methods still cause misrecognition. One of the causes of this misrecognition I
: The system unifies the IDs (recited word sounds) of each speaker used for verification, and as a result, for some people,
The probability that the word or word string of that ID contains personal characteristics is small (there have been cases where the matching device has not been able to complete the matching. For example, if the system specifies the ID as "O (zero)", Person C2 always pronounces "Zero" in a stable tone, so it has a good sense of individuality (stability) and is easy to identify as an individual, but for another person, "Zero" is pronounced in a stable tone.
"Zero" is uttered in an unstable tone each time, so it does not contain much individuality and is difficult to identify.

Even if the user's favorite ID is uttered, even if it includes individuality, the matching device must match various IDs (word recognition, voice feature matching), so it is compatible with all speakers. The drawback was that it couldn't be done.

[Purpose of the invention]

It is an object of the present invention to provide a speaker verification method in which a device can generate an ID that is easy to verify by itself, thereby improving the verification rate.

[Summary of the invention]

The present invention provides a means for creating a dictionary for one-speaker verification (2) a means for analyzing and recording individual voice characteristics for multiple words uttered by a registrant, and a step for registering these voice characteristics. For each word (equipped with a means to perform two-speaker matching, IDs for each individual's matching are created using the phonetic characteristics of words with good matching results)
In the speaker verification method that has a means for determining the ID, when performing verification, the ID is used and verification is performed by two means: word recognition of this ID and speaker recognition based on the utterance of the speaker. This is a characteristic feature.

〔Effect of the invention〕

According to the present invention, it is possible to reduce errors due to verification (2) by giving an ID with highly individual (highly stable) voice characteristics to each registrant ζ2, which improves security. Practicality is also improved from this point of view.

[Embodiments of the invention]

Embodiments of the present invention will be described below with reference to the drawings. FIG. 1 is a flow chart of a first embodiment of the present invention, and FIG. 2 is a block diagram of the first embodiment. In the first embodiment, when a registrant creates a dictionary, [2] the registrant utters a plurality of words specified by the system in advance, detects their voice characteristics, registers them in a temporary dictionary, and repeats them a certain number of times. When this occurs (after at least one utterance has been completed), the registration in the temporary dictionary is stopped, and the registrant is asked to utter the plurality of words in sequence again, thereby detecting the voice characteristics of each word.

The ID-Q is determined by comparing these with the voice features registered in the temporary dictionary, and the formal registration is made. First, the user must either check the speaker or create a dictionary to request registration. selects all functions (1st N step 11,
12). That is, in the function selection section 102 of FIG. 2, a switch is used to select whether to perform speaker verification or registration. At this time.

If registration is selected, counter 10 is pressed to create a dictionary.
3 is initialized (step 13 in FIG. 1) and switch C is selected. Although it is not shown in Figure 2, in the initial settings, the C double line is uttered repeatedly for registration and verification, so the counter that counts the number of times it is uttered must be cleared (N = 0), all voices. Set the number of inputs M. For example, the numbers "O (zero)", "l (ichi) j, r2 (ni)"...
...If you ask someone to say "9 (kyu)" twice, M = 20. After performing these initial settings, the second
In the system shown in the figure, the user is requested to input the sound of each word (two lines in sequence) via a display or the like (not shown) (for example, the user is requested to input "zero" first).

Step 14 in Figure 1) When the user utters a voice ("zero") in accordance with the word input request (Step 15 in Figure 1)
This input voice is subjected to conversion, spectrum analysis, etc. in the analysis unit 1011, and is converted into a series of characteristic parameters (step 16 in FIG. 1). This analysis part lot is 1. For example, if the input voice is "zero" (/ze//ro/), the overall feature (parameter and its vowel part (/xl
) to detect both feature parameters. The detected overall feature parameters and vowel feature parameters are registered in the provisional memory 104 via switches B and C (step 18 in FIG. 1). The number of voice inputs is incremented by one (
Step 19 in FIG. 1) A request for input of the first word voice (for example, "ichi") is made (Step 20 in FIG. 1).

In this way, registration is performed in the temporary dictionary memory 104 until the counter 103 reaches M/2 (half of the total number of voice inputs).
m1 diagram step 17), on M/21 (when it becomes -, the switch is set to the D side, and the speaker collation unit 105 uses the temporary memory that has been used for registration until now to match the newly input voice C). (si diagram step 21)
゜For this newly inputted voice C2, the analysis unit lO1 detects only the parameters of the vowel part from V2 or higher (indication of the counter 103 which is lower than 2).

For example, up to the 10th time, the user will be asked "0 (zero) J -, 4
・The vowel parameters (and overall parameters) are registered by uttering ``l (ichi)'', ......``9 (gyu)'', and from the 11th to the 20th time, the user (2 ``
0 (zero) J, rl (ichi)", ... "9 (
``Kyuu)'' and select each vowel part parameter (r
! /Kobun Raba/Ku/, if it is "Ite", /i/) are sequentially compared with the already registered vowel part parameters. This speaker verification is performed using, for example, similarity calculation or distance calculation. These matching results are sent to the C2 discriminator 106 together with the vowel partial parameters and overall feature parameters that were used for matching (registered in the temporary dictionary).

tts1 Figure Stella Step 22 during dictionary registration C2 speaker verification and transfer to the discrimination unit 106 is performed until N=M (Figure 1 Steps 23, 24, 25)
.

If the number of voice inputs N becomes N-Mζ2, the determination unit 106 determines t' from the instruction ζ2 of the counter 103! ,, Select the phonetic features (vowels and overall parameters) of the word with the most correct matching result (high degree of similarity) among the results of speaker matching (that is, from "zero" to "kyu") If "zero" has the highest degree of similarity, then [zero J
+2 included vowel part and overall feature parameters) I
The data is output to the D creation unit 107 (step 26 in FIG. 1). I
In response to this result, the D creation unit 107 creates an ID (step 27 in FIG. 1, for example, sets the ID to "zero" from the received overall feature parameters), and this dictionary memory 1.
08 (step 28 in Figure 1), and outputs the ID to the user via a display or the like (step 29 in Figure m1). Here, as the format of the ID registered in the main dictionary memory 108, 1. For example, if the ID is "O (zero)", as mentioned above (2. The parameters of the vowel part are stored as a pair.

On the other hand, when performing speaker verification using the ID created by the above method ζ2, the switch of the function selection section 102 is turned on by the user's instruction C2 [Step 11 in FIG.
, 12). Next, when the second user utters the ID that he or she has memorized (Step 30 in Figure 1), this input voice is converted into feature parameters for the whole and vowel parts in the Yu analysis section 101 as described above. Step 31 in Figure 1).The word recognition unit 110 inputs the feature parameters of the entire word uttered by the user as an ID, and determines if this matches the feature parameters of the entire word registered in advance in the dictionary memo 9-104. It recognizes (verifies) whether or not there is a
Step 32 in FIG. 1), the speaker verification unit 109 verifies whether the characteristic parameters of the vowel part of the input speech match the characteristic parameters of the vowel part registered as an ID (Step 33 in Figure 1). ). The matching unit ill uses these recognition results and matching results to find the final speaker (
ID registrant) is verified (step 34 in FIG. 1), and the result is output (step 35 in FIG. 1).

According to the above-mentioned Example C2, the 1-speaker matching system selects in advance from among a plurality of words that are easy for the system itself to match (recognize) the words that are the most personal (stable) and easy to match for each speaker's text 2. Since an ID is created, it is possible to improve the accuracy rate of verification.

Next, a second embodiment of the present invention will be explained with reference to the drawings. FIG. 3 shows the front yard of the second embodiment, and the fourth
The figure is a block diagram of the second embodiment. In this embodiment, when creating a speaker ID (2), in the dictionary registration stage as described above, the entire feature parameters and vowel part parameters of the input speech are registered in the temporary dictionary, and when repeating the input speech, word recognition (entire This method performs speaker verification (vowel feature parameter verification) and speaker verification (vowel feature parameter verification), creates an ID from these two results, and registers it in this dictionary.

@31J, in FIG. 4, the difference from FIGS. 1 and 2 is the word recognition unit 112. Word recognition (Step 36 in Figure 1)
This is the place where it is mentioned.

When dictionary registration is selected as described above, the voice input by the speaker is converted into the overall feature parameter and the vowel part feature parameter (two conversions) in the analysis unit 101, and this pair is used as a temporary dictionary memo!J 104 will be registered to.

When this registration is repeated up to M/2 times (when voice input for all words is completed), the switch is switched to the D side according to instructions from the counter 103, and the following input voice (all words (voices repeated with y) ) is compared by the speaker matching unit 105 and the word recognition unit 112.In other words, in the old version of analysis flbt, the word voice (2) that has been recursed is detected by detecting the overall feature parameter and the feature parameter of the vowel part. The latter is sent to the word recognition section 112, and the latter is sent to the speaker verification section [05.

The word recognition unit 112 compares the sent overall feature parameters with the overall feature parameters registered in advance in the temporary dictionary memo 9-104, and sends the matching results together with the overall parameters to the semi-separation unit 106 (the Step 36 in Figure 3].The speaker matching unit 105 matches the sent vowel part parameters with the vowel part parameters registered in the temporary dictionary memo!J-104, and sends the matching results together with the vowel part parameters to the discrimination unit 106. (Step 21 in FIG. 3). After this process is repeated M times, according to the instruction from the counter 103, the discriminator 106 selects the overall parameter and the vowel partial parameter if both matching results are good. It is sent to the ID creation unit 107 (@Steps 22 and 26 in Figure 3).Here, depending on the case, it is possible to weight the two matching results for discrimination and select parameters.In this way, the L
The D creation capital 107 creates an ID using the selected overall parameter and its vowel parameter, and registers these two parameters in the main dictionary memory 108 as the ID. (3rd
Figure Step 27.28. For example, if the selected parameter is (
/ze//ro/) and (/工/), the ID is "zero"
(These two parameters become the ID parameters.) According to the second embodiment, the speaker verification system uses the speaker's input sound yH! In contrast, it has the highest individuality (vowel parameter matching rate), and word recognition (overall parameter matching)
Since an ID with the highest recognition rate is created, the speaker verification rate can be further improved, and problems with security i2 are also reduced.

Note that the present invention is not limited to the above embodiments. For example, when creating an ID, if there is a part other than the required verification part in the ID, the registrant may create this part. Furthermore, various methods known from Japan and the United States may be suitably employed as methods for detecting feature parameters of input speech and for collation (recognition) processing.

In short, the present invention can be implemented with various modifications without departing from the gist thereof.

[Brief explanation of the drawing]

Figure 1 is a flowchart of the @l embodiment of the present invention, Figure 2 is the first block diagram of the present invention, Figure 3 is a flowchart of the second embodiment of the present invention, and Figure 4 is the flowchart of the present invention. FIG. 3 is a block diagram of a second embodiment of the invention. 101...Analysis unit 102...Function selection unit 103...Counter 104...Temporary oath memory 105...Speaker verification unit 106...Discrimination unit 107...ID creation capital 108...・Book dictionary memory 109...Speaker collation unit 110...Word recognition unit 111...Collation unit 112..." Vocabulary magazine Oribe agent Patent attorney Nori Chika Keisuke (and 1 other person)

Claims (3)

[Claims] (1) A detection means for detecting each voice feature of a plurality of word sounds uttered by a speaker; a storage means for storing each voice feature detected by the detection means; collation means for collating each voice feature re-detected by the detection means for the plurality of word sounds and each voice feature stored by the storage means;
and determining means for determining the ID of the speaker based on the verification result by the verification means. (2) The collation means recognizes the plurality of word sounds uttered again by the speaker and collates the voice characteristics of each, and the determination means determines an ID based on the recognition result and the collation result by the collation means. A speaker verification method according to claim 1, characterized in that: (3) The detecting means detects voice features from the word sounds uttered by the speaker, and the matching means determines whether or not the voice features match the voice features of the ID determined by the determining means. A speaker verification method according to claim 1, characterized in that verification is performed.