CA1207907A - Speaker verification system - Google Patents
Speaker verification systemInfo
- Publication number
- CA1207907A CA1207907A CA000450839A CA450839A CA1207907A CA 1207907 A CA1207907 A CA 1207907A CA 000450839 A CA000450839 A CA 000450839A CA 450839 A CA450839 A CA 450839A CA 1207907 A CA1207907 A CA 1207907A
- Authority
- CA
- Canada
- Prior art keywords
- pattern
- verification
- control reference
- patterns
- dissimilarities
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Expired
Links
Classifications
-
- G—PHYSICS
- G10—MUSICAL INSTRUMENTS; ACOUSTICS
- G10L—SPEECH ANALYSIS OR SYNTHESIS; SPEECH RECOGNITION; SPEECH OR VOICE PROCESSING; SPEECH OR AUDIO CODING OR DECODING
- G10L17/00—Speaker identification or verification
Abstract
ABSTRACT
A speaker verification system includes an input analyzer for analyzing a word spoken by a registered speaker or a person being verified, and outputting the result as an input pattern. A
registration number designator is provided for outputting a regis-tration number and an input pattern buffer is provided for storing the input pattern. A verification reference pattern memory stores the input pattern as a verification reference pattern during regis-tration, designated by the registration number. A control refer-ence pattern memory stores, as control reference patterns, patterns which are obtained when several predetermined persons speak a predetermined password. A pattern matching unit is provided for determining first dissimilarities between the verification refer-ence pattern and the control reference patterns, a second dissimi-larity between the input pattern of the person being verified, which is obtained during the verification, and the verification reference pattern, and third dissimilarities between the input pattern and the control reference patterns. A control reference pattern designator is provided for selecting several first dissimi-larities consecutively, starting from the smallest of the first dissimilarities, to designate the control reference patterns cor-responding to the selected dissimilarities as the control reference patterns corresponding to the registration number. Finally, a judgement unit is provided for generating the confirmation signal when the second dissimilarity is less than both the third dissimi-larities corresponding to the designated control reference patterns and a predetermined threshold value.
A speaker verification system includes an input analyzer for analyzing a word spoken by a registered speaker or a person being verified, and outputting the result as an input pattern. A
registration number designator is provided for outputting a regis-tration number and an input pattern buffer is provided for storing the input pattern. A verification reference pattern memory stores the input pattern as a verification reference pattern during regis-tration, designated by the registration number. A control refer-ence pattern memory stores, as control reference patterns, patterns which are obtained when several predetermined persons speak a predetermined password. A pattern matching unit is provided for determining first dissimilarities between the verification refer-ence pattern and the control reference patterns, a second dissimi-larity between the input pattern of the person being verified, which is obtained during the verification, and the verification reference pattern, and third dissimilarities between the input pattern and the control reference patterns. A control reference pattern designator is provided for selecting several first dissimi-larities consecutively, starting from the smallest of the first dissimilarities, to designate the control reference patterns cor-responding to the selected dissimilarities as the control reference patterns corresponding to the registration number. Finally, a judgement unit is provided for generating the confirmation signal when the second dissimilarity is less than both the third dissimi-larities corresponding to the designated control reference patterns and a predetermined threshold value.
Description
1~'07~9~7 Speaker Verification System BACKGROUND OF THE INVENTION
The present invention relates to a speaker vertification system.
Systems in which a spoken password is compared with a password which has been registered in advance, to determine whether the speaker is the registered person, are called "speaker verification systems". It is thought that these systems will be useful as identification-verifying mea~s in banks or check-in and check-out control systems, and practical realization thereof is expected.
A conventionally developed speaker verification system is practised in the following manner. In this system, the voice pattern o~ a password spoken by a registered speaker is stored as a reference pattern A and, during verification, an input password voice pattern X spoken by the speaker is compared with the reference pattern and the dissimilarity D (X, A) therebetween is calculated. If the dissimilarity is less than a predetermined threshold value ~, it is recognized that the input voice pattern coincides with the reference pattern, and the speaker is judged to be the registered person. Conversely, if the dissimilarity is greater than the threshold value ~, it is recognized that the two patterns do not coincide, and the speaker is judged to be an impostor.
~r~
The reliability of the prior art voice recognition system which has the function of verifying the speaker as described above is greatly influenced by the magnitude of the threshold value 0. There are two different kinds of erroneous operation of the speaker verification system, i.e., a false rejection of the true registered speaker as an impostor, and a false acceptance of an impostor as the true registered speaker. When the threshold value ~ is reduced, the possibility of false rejections can be reduced, but the possibility of false acceptances is increased.
Conversely, if the threshold value ~ is increased, false rejections can be reduced, but false acceptances are increased.
When applied to person verification in banks or check-in and check-out systems, it is thougntthat the threshold value ~ should be set to be rather low, in view of the seriousness of the result that might occur from a false acceptance. This inevitably leads to a system in which, when there is a false rejection, the speaker must pronounce the password repeatedly until it is correctly accepted. This system, however, leads to a reduction in service with person verification in banks, or to the inconvenience of blocked doorways with check-in and check-out system~. Moreover, it is difficult to set a suitable threshold value in this system, therefore, the following problems are caused. Even when the same person .
.
'~2~7!~7 pronounces the predetermined password, the voice pattern changes each time, and the degree of change is different for each person.
If the threshold value is set at a strict (or low) value, there-fore, it often happens that a true registered speaker is recog-nized as an imposter. Conversely, of the threshold value is set at a rather loose (or high) level, it follows that an imposter could be erroneously recognized as a true registered speaker.
A system which has improved on these problems has been proposed by us, the inventors (in Japanese Patent Disclosure 58-76893 published on ~ay 10, 1983). According to this proposed system, not only the reference pattern of a password, but also voice patterns which are similar to but different from the pass-word are registered as control reference patterns, and the speaker is recognized to be the registered person, and a coincidence signal is generated, only when the distance obtained as a result of matching with an input pattern satisfies the following condi-tions:
(a) The distance from the reference pattern is less than the distance from the control reference patterns; and (b) The distance from the reference pattern is less than a predetermined threshold value.
According to this system, because of ]udgement standard (a), the threshold value used in standard (b) can be loosely set. This provides the effects that false recognitions of true registered speakers as impostors are reduced, and that , :~2~7~907 false accePtances of impostors as true registered speakers , can be prevented by standard (a).
This system, however, has the ~roblem that control reference ~atterns must be re~istered by each nerson.
S Esnecially in the case of the ~eneral ~ublic use, this ~!70uld increase false or erroneous o~erations due to the arbitrarily selection and registration by each Person for the control reference Patterns.
SUMMARY OF THE INVENTION
It is, therefore, an object of the ~resent invention to provide a speaker verification system which is capable of reducin~ erroneous recognitions and registration work.
Another object of the present invention is to provide a speaker verification system which eliminates the problems in operation.
r~ A further ob.~ect of the present invention is to provide a speaker verification system which is extremely secure and safe.
Accordin~ to a feature o.f the present invention, a s~eaker verification system is provided comprising, a first memory for storing, as control reference patterns, patterns which are obtained by s~eaking a predetermined password under a plurality of predetermined conditions, a second memory for storing, as a verification reference pattern corresnonding to a re~istered s~eaker, a pattern 7~
which is obtained by the utterance of the password by the registered speaker, means for selecting a predetermined number of control reference patterns from the first memory consecutivel~t, from those which are the most similar to the verificati~n reference pattern corresponding to the registered speaker, means for determining, during verification, a first dissimilarity between the input pattern obtained by the speech of the person being verified and the verification reference pattern corresponding to the registered speaker, and second dissimilarities between the input pattern and the selected control reference patterns, and means for outputting a confirma~ion signal when the first dissimilarity is less than both the second dissimilarities and a predetermined threshold value.
BRIEF DESCRIPTION OF THE DRAWINGS
-Fig. 1 is a block diagram of an embodiment of the speaker verification system according to the present invention; and Fig. 2 is a block diagram of another embodiment of the present invention.
DESCRIPTION OE~ THE PREFERRED EMBODIMENTS
.. . .. ~
The present invention uses as a control reference patterns obtained when a plurality of predetermined persons speak an identical password, in this way it . , ~L~Q~
differs from the previous Japanese Patent Disclosure in which patterns obtained by the utterance of words similar in pronunci-ation to a verification reference pattern are used for each registered speaker. As a result, the registration of the control reference patterns is done once , and no operational problems occur.
Referring to Fig. 1, a control reference pattern memory 100 holds password voice patterns of a large number of people. For example, the memory lO0 holds the patterns of the word "Mississippi (with the pronunciation /misasipi/) spoken by l,000 people. The set of the control reference patterns is designated by:
[Bl B2 _ _ _, Bm, - - -, and B ]-Each control reference pattern Bm is expressed, as in the previous Japanese Patent Disclosure, by a time series of charac-teristic vectors.
The operation of the present system is divided into two modes, registration and verification. A switch 50 is trans-ferred to a terminal 51 in registration mode and to a terminal 52 in verification mode.
First of all, the operation relating to the registration of the verification reference pattern will be described. A
registration number-designating unit 20, such as a keyboard for example, is used to input data on the person being registered, i.e., a registration number _. Subsequently, the password /misasipi/ is input via a 1~7~
microphone 10. This voice is analyzed by an analyzer 30 and is then input to an input pattern buffer 40 (in which the input will be referred to as "input pattern Xi').
The input pattern X is designated verification reference pattern An by the registration number n and is supplied to a verification reference pattern memory 60 through the switch 50.
After that, a counter 110 varies a control reference pattern designation signal _ consecutively from 1 to 1000.
From the control reference pattern memory 100, accordingly, the control reference patterns B are output in sequence from sl to B1000 and are supplied to a pattern matching unit 70.
The pattern matching unit 70 compares the input pattern X and the control reference patterns Bm from the control reference pattern memory 100, and consecutively calculates distances D(~, Bm). The distances DtX, Bm) are consecutively supplied to a control designation and decision unit 80 in which they are compared with one another so that the shortest distance to, for example, the fifth shortest distance are determined, and the corresponding numbers m axe sent to a control designator 90.
The numbers m are assigned in the control designator 90 to the registration numbers n and are stored in the following form:
m(n, 1), m(n, 2), - - - m(n, 5) (2).
~o~
Above-stated processes provide the verification . r~ference patterns An, and prepare for the control designation, as in expression (2). The processing thus far described is repeated, changing the registration number n, every time a new person is to be registered.
Next the portions of the diagram relating to the verification will be described. A person re~uiring verification operates the keys of the keyboard 20 to input his registration number _ or registration data, and utters the password/mis~sipi/ through the microphone 10.
This password utterance is analyzed by the analyzer 30 in the same way as in registration, and is input as the input pattern X to the input pattern buffer 40.
When the registration number n is input to the verification reference pattern memory 60, the verification reference pattern A is accordingly output. In response to this verification reference pattern An, the pattern matching unit 70 first calculates the distance D(X, An).
Responsive to the registration number n, the control designator 90 outputs a control designation signal in the form of expression (2). In response to this control designation signal, the control reference pattern memory 100 generates the following control reference patterns:
Bm: m = m(n, 1), m(n, 2), - - -, m(n, 5) (3) The pattern matching unit 70 compares the reference patterns Bm and the input pattern X sent through the switch 50 and calculates the following distances:
:
:IIL'~07!~7 D(X, Bm): m = m(n, l), m(n, 2), - - -, m(n, 5) (4) A judgement unit 120 compares the distances D(X, Am) and the group of distances in the form of expression (4) to determine that the person being verified is the true registered speaker only if ~
D(X, A ) < D(X, B ), : m = m(n, 1), m(n, 2), - - -, m(n, 5), and t2~ : D(X, An) ~ a predetermined threshold value.
According to the construction thus far described, ,-i 10 effects similar to those of the previous ~. Patent ~; ~c_\o~
~ppli~i~n can be obtained, but each person being verified does not need to speak and register control reference patterns.
The above description directed to the fundamental construction of the present invention uses the case in which distance (or dissimilarity) is used for the comparison of voice patterns. Ho~ever, a measure such as a correlation, which has inverse magnitude relation with the distance, may be employed. In this modification, the inequality signs of judgement standards tl~ and L 2 are naturally inverted. It is also possible to conduct the judgement by resorting to judgement standard ~l~ alone.
This judgement is equivalent to the case in which the threshold value of standard (2~ is set at a sufficiently high value.
Fig. 2 is a block diagram of another embodiment of the present invention. This embodiment can cope with the piracy of the password by using the technique of the previous Japanese Patent Disclosure in combination with the present invention.
In Fig. 2, reference numerals identical to those of Fig. 1 indicate basically identical components. The judgement unit 120 not only has the judgement function of the embodiment of Fig. l, but also the judgement function of the previous Japanese Patent Disclosure. Moreover, a control reference pattern memory 130 holds the patterns words such as: /mes a sipi/, mis a sépi/
or /musasipi/, that have a pronunciation similar to that of the password /misasipi/~ designated by Anl, A2n and A3n, respectively, and the control reference patterns Anl, A2n and An3 of the registered speaker corresponding to the registration number _ are supplied to the pattern matching unit 70.
In verification mode, the switch 50 is transferred to the terminal 52. When a person claiming to be the true registered speaker operates the keyboard 20 and speaks the password, the output from the analyzer 30 is supplied to the input pattern buffer 40. In the pattern matching unit 70, the input pattern X from the switch 50 and the corresponding reference patterns An, Anl, A2n and A3n for the registration number _, supplied from the ~:~0~ 07 verification reference pattern memory 60 and the control reference pattern memory 130, are compared, and the distances D(X, A ), D(X, Aln), D(X, A2n) and D(X, A3) are calculated and then supplied to the judgement unit 120.
Like in the previous embodiment of Fig. 1, the judgement unit 120 uses the group of distances D (X, Bm) of expression (4) obtained in the pattern matching unit 70 to judge that a first verification has been conducted, depending upon whether conditions ~1~ and ~2~ are satisfied or not, i.e., when D(X, An) is less than all of D (X, Bm) and the predetermined threshold value ~1 The minimum distance Dmin of the distances thus obtained is determined in the judgement unit 120, and it is judged that a second verification has been conducted when the reference pattern corresponding to the minimum distance Dmin is An and the minimum distance Dmin less than a threshold value ~2~ namely, D(X, Am) is less than all of D(X, Aln), D(X, A2n), D(X, An3) and the value ~2. When the reference pattern corresponding to Dmin is either of Anl, A2n and A3n, or when the minimum distance Dmin is more than the threshold value ~2' namely, D(X, An) is larger than either of D(X, Al), D(X, A2), D(X, A3) or the value 2' on the other hand, it is judged that the second verification is not conducted. The judgement unit 120 outputs a determination signal r indicating that the person being verified is the true registered speaker, 07~7 when the first and second verifications have been conducted.
r, The second embodiment has a more tedious registration procedure because it uses control patterns for words similar to the password. Since, however, the final judgement is conducted by a logical product of first and second verifications, its security and safety are remarkably improved. For example, if the password for the second verification is a secret one which is different for each registered person, and if the password used for the first verification is shared by the registered people, the possibility of a false acceptance is greatly reduced, even if the common password is stolen, because there is the barrier of the secret password. From the above description, it is apparent that different passwords or the same password can be used as two passwords for the first and second verifications.
The present invention relates to a speaker vertification system.
Systems in which a spoken password is compared with a password which has been registered in advance, to determine whether the speaker is the registered person, are called "speaker verification systems". It is thought that these systems will be useful as identification-verifying mea~s in banks or check-in and check-out control systems, and practical realization thereof is expected.
A conventionally developed speaker verification system is practised in the following manner. In this system, the voice pattern o~ a password spoken by a registered speaker is stored as a reference pattern A and, during verification, an input password voice pattern X spoken by the speaker is compared with the reference pattern and the dissimilarity D (X, A) therebetween is calculated. If the dissimilarity is less than a predetermined threshold value ~, it is recognized that the input voice pattern coincides with the reference pattern, and the speaker is judged to be the registered person. Conversely, if the dissimilarity is greater than the threshold value ~, it is recognized that the two patterns do not coincide, and the speaker is judged to be an impostor.
~r~
The reliability of the prior art voice recognition system which has the function of verifying the speaker as described above is greatly influenced by the magnitude of the threshold value 0. There are two different kinds of erroneous operation of the speaker verification system, i.e., a false rejection of the true registered speaker as an impostor, and a false acceptance of an impostor as the true registered speaker. When the threshold value ~ is reduced, the possibility of false rejections can be reduced, but the possibility of false acceptances is increased.
Conversely, if the threshold value ~ is increased, false rejections can be reduced, but false acceptances are increased.
When applied to person verification in banks or check-in and check-out systems, it is thougntthat the threshold value ~ should be set to be rather low, in view of the seriousness of the result that might occur from a false acceptance. This inevitably leads to a system in which, when there is a false rejection, the speaker must pronounce the password repeatedly until it is correctly accepted. This system, however, leads to a reduction in service with person verification in banks, or to the inconvenience of blocked doorways with check-in and check-out system~. Moreover, it is difficult to set a suitable threshold value in this system, therefore, the following problems are caused. Even when the same person .
.
'~2~7!~7 pronounces the predetermined password, the voice pattern changes each time, and the degree of change is different for each person.
If the threshold value is set at a strict (or low) value, there-fore, it often happens that a true registered speaker is recog-nized as an imposter. Conversely, of the threshold value is set at a rather loose (or high) level, it follows that an imposter could be erroneously recognized as a true registered speaker.
A system which has improved on these problems has been proposed by us, the inventors (in Japanese Patent Disclosure 58-76893 published on ~ay 10, 1983). According to this proposed system, not only the reference pattern of a password, but also voice patterns which are similar to but different from the pass-word are registered as control reference patterns, and the speaker is recognized to be the registered person, and a coincidence signal is generated, only when the distance obtained as a result of matching with an input pattern satisfies the following condi-tions:
(a) The distance from the reference pattern is less than the distance from the control reference patterns; and (b) The distance from the reference pattern is less than a predetermined threshold value.
According to this system, because of ]udgement standard (a), the threshold value used in standard (b) can be loosely set. This provides the effects that false recognitions of true registered speakers as impostors are reduced, and that , :~2~7~907 false accePtances of impostors as true registered speakers , can be prevented by standard (a).
This system, however, has the ~roblem that control reference ~atterns must be re~istered by each nerson.
S Esnecially in the case of the ~eneral ~ublic use, this ~!70uld increase false or erroneous o~erations due to the arbitrarily selection and registration by each Person for the control reference Patterns.
SUMMARY OF THE INVENTION
It is, therefore, an object of the ~resent invention to provide a speaker verification system which is capable of reducin~ erroneous recognitions and registration work.
Another object of the present invention is to provide a speaker verification system which eliminates the problems in operation.
r~ A further ob.~ect of the present invention is to provide a speaker verification system which is extremely secure and safe.
Accordin~ to a feature o.f the present invention, a s~eaker verification system is provided comprising, a first memory for storing, as control reference patterns, patterns which are obtained by s~eaking a predetermined password under a plurality of predetermined conditions, a second memory for storing, as a verification reference pattern corresnonding to a re~istered s~eaker, a pattern 7~
which is obtained by the utterance of the password by the registered speaker, means for selecting a predetermined number of control reference patterns from the first memory consecutivel~t, from those which are the most similar to the verificati~n reference pattern corresponding to the registered speaker, means for determining, during verification, a first dissimilarity between the input pattern obtained by the speech of the person being verified and the verification reference pattern corresponding to the registered speaker, and second dissimilarities between the input pattern and the selected control reference patterns, and means for outputting a confirma~ion signal when the first dissimilarity is less than both the second dissimilarities and a predetermined threshold value.
BRIEF DESCRIPTION OF THE DRAWINGS
-Fig. 1 is a block diagram of an embodiment of the speaker verification system according to the present invention; and Fig. 2 is a block diagram of another embodiment of the present invention.
DESCRIPTION OE~ THE PREFERRED EMBODIMENTS
.. . .. ~
The present invention uses as a control reference patterns obtained when a plurality of predetermined persons speak an identical password, in this way it . , ~L~Q~
differs from the previous Japanese Patent Disclosure in which patterns obtained by the utterance of words similar in pronunci-ation to a verification reference pattern are used for each registered speaker. As a result, the registration of the control reference patterns is done once , and no operational problems occur.
Referring to Fig. 1, a control reference pattern memory 100 holds password voice patterns of a large number of people. For example, the memory lO0 holds the patterns of the word "Mississippi (with the pronunciation /misasipi/) spoken by l,000 people. The set of the control reference patterns is designated by:
[Bl B2 _ _ _, Bm, - - -, and B ]-Each control reference pattern Bm is expressed, as in the previous Japanese Patent Disclosure, by a time series of charac-teristic vectors.
The operation of the present system is divided into two modes, registration and verification. A switch 50 is trans-ferred to a terminal 51 in registration mode and to a terminal 52 in verification mode.
First of all, the operation relating to the registration of the verification reference pattern will be described. A
registration number-designating unit 20, such as a keyboard for example, is used to input data on the person being registered, i.e., a registration number _. Subsequently, the password /misasipi/ is input via a 1~7~
microphone 10. This voice is analyzed by an analyzer 30 and is then input to an input pattern buffer 40 (in which the input will be referred to as "input pattern Xi').
The input pattern X is designated verification reference pattern An by the registration number n and is supplied to a verification reference pattern memory 60 through the switch 50.
After that, a counter 110 varies a control reference pattern designation signal _ consecutively from 1 to 1000.
From the control reference pattern memory 100, accordingly, the control reference patterns B are output in sequence from sl to B1000 and are supplied to a pattern matching unit 70.
The pattern matching unit 70 compares the input pattern X and the control reference patterns Bm from the control reference pattern memory 100, and consecutively calculates distances D(~, Bm). The distances DtX, Bm) are consecutively supplied to a control designation and decision unit 80 in which they are compared with one another so that the shortest distance to, for example, the fifth shortest distance are determined, and the corresponding numbers m axe sent to a control designator 90.
The numbers m are assigned in the control designator 90 to the registration numbers n and are stored in the following form:
m(n, 1), m(n, 2), - - - m(n, 5) (2).
~o~
Above-stated processes provide the verification . r~ference patterns An, and prepare for the control designation, as in expression (2). The processing thus far described is repeated, changing the registration number n, every time a new person is to be registered.
Next the portions of the diagram relating to the verification will be described. A person re~uiring verification operates the keys of the keyboard 20 to input his registration number _ or registration data, and utters the password/mis~sipi/ through the microphone 10.
This password utterance is analyzed by the analyzer 30 in the same way as in registration, and is input as the input pattern X to the input pattern buffer 40.
When the registration number n is input to the verification reference pattern memory 60, the verification reference pattern A is accordingly output. In response to this verification reference pattern An, the pattern matching unit 70 first calculates the distance D(X, An).
Responsive to the registration number n, the control designator 90 outputs a control designation signal in the form of expression (2). In response to this control designation signal, the control reference pattern memory 100 generates the following control reference patterns:
Bm: m = m(n, 1), m(n, 2), - - -, m(n, 5) (3) The pattern matching unit 70 compares the reference patterns Bm and the input pattern X sent through the switch 50 and calculates the following distances:
:
:IIL'~07!~7 D(X, Bm): m = m(n, l), m(n, 2), - - -, m(n, 5) (4) A judgement unit 120 compares the distances D(X, Am) and the group of distances in the form of expression (4) to determine that the person being verified is the true registered speaker only if ~
D(X, A ) < D(X, B ), : m = m(n, 1), m(n, 2), - - -, m(n, 5), and t2~ : D(X, An) ~ a predetermined threshold value.
According to the construction thus far described, ,-i 10 effects similar to those of the previous ~. Patent ~; ~c_\o~
~ppli~i~n can be obtained, but each person being verified does not need to speak and register control reference patterns.
The above description directed to the fundamental construction of the present invention uses the case in which distance (or dissimilarity) is used for the comparison of voice patterns. Ho~ever, a measure such as a correlation, which has inverse magnitude relation with the distance, may be employed. In this modification, the inequality signs of judgement standards tl~ and L 2 are naturally inverted. It is also possible to conduct the judgement by resorting to judgement standard ~l~ alone.
This judgement is equivalent to the case in which the threshold value of standard (2~ is set at a sufficiently high value.
Fig. 2 is a block diagram of another embodiment of the present invention. This embodiment can cope with the piracy of the password by using the technique of the previous Japanese Patent Disclosure in combination with the present invention.
In Fig. 2, reference numerals identical to those of Fig. 1 indicate basically identical components. The judgement unit 120 not only has the judgement function of the embodiment of Fig. l, but also the judgement function of the previous Japanese Patent Disclosure. Moreover, a control reference pattern memory 130 holds the patterns words such as: /mes a sipi/, mis a sépi/
or /musasipi/, that have a pronunciation similar to that of the password /misasipi/~ designated by Anl, A2n and A3n, respectively, and the control reference patterns Anl, A2n and An3 of the registered speaker corresponding to the registration number _ are supplied to the pattern matching unit 70.
In verification mode, the switch 50 is transferred to the terminal 52. When a person claiming to be the true registered speaker operates the keyboard 20 and speaks the password, the output from the analyzer 30 is supplied to the input pattern buffer 40. In the pattern matching unit 70, the input pattern X from the switch 50 and the corresponding reference patterns An, Anl, A2n and A3n for the registration number _, supplied from the ~:~0~ 07 verification reference pattern memory 60 and the control reference pattern memory 130, are compared, and the distances D(X, A ), D(X, Aln), D(X, A2n) and D(X, A3) are calculated and then supplied to the judgement unit 120.
Like in the previous embodiment of Fig. 1, the judgement unit 120 uses the group of distances D (X, Bm) of expression (4) obtained in the pattern matching unit 70 to judge that a first verification has been conducted, depending upon whether conditions ~1~ and ~2~ are satisfied or not, i.e., when D(X, An) is less than all of D (X, Bm) and the predetermined threshold value ~1 The minimum distance Dmin of the distances thus obtained is determined in the judgement unit 120, and it is judged that a second verification has been conducted when the reference pattern corresponding to the minimum distance Dmin is An and the minimum distance Dmin less than a threshold value ~2~ namely, D(X, Am) is less than all of D(X, Aln), D(X, A2n), D(X, An3) and the value ~2. When the reference pattern corresponding to Dmin is either of Anl, A2n and A3n, or when the minimum distance Dmin is more than the threshold value ~2' namely, D(X, An) is larger than either of D(X, Al), D(X, A2), D(X, A3) or the value 2' on the other hand, it is judged that the second verification is not conducted. The judgement unit 120 outputs a determination signal r indicating that the person being verified is the true registered speaker, 07~7 when the first and second verifications have been conducted.
r, The second embodiment has a more tedious registration procedure because it uses control patterns for words similar to the password. Since, however, the final judgement is conducted by a logical product of first and second verifications, its security and safety are remarkably improved. For example, if the password for the second verification is a secret one which is different for each registered person, and if the password used for the first verification is shared by the registered people, the possibility of a false acceptance is greatly reduced, even if the common password is stolen, because there is the barrier of the secret password. From the above description, it is apparent that different passwords or the same password can be used as two passwords for the first and second verifications.
Claims (8)
1. A speaker verification system comprising:
a first memory for storing, as control reference patterns, patterns which are obtained by speaking a predetermined password under a plurality of predetermined conditions;
a second memory for storing, as a verification reference pattern corresponding to a registered speaker, a pattern which is obtained by said registered speaker speaking said password;
means for selecting a predetermined number of the control reference patterns from said first memory consecutively, starting from said control reference pattern which is the most similar to the verification reference pattern corresponding to said registered speaker;
means for determining, during verification, a first dissimilarity between the input pattern obtained by the speech of a person being verified and the verification reference pattern corresponding to said registered speaker, and second dissimilarities between said input pattern and the selected control reference patterns; and means for outputting a confirmation signal when said first dissimilarity is less than both said second dissimilarities and a predetermined threshold value.
a first memory for storing, as control reference patterns, patterns which are obtained by speaking a predetermined password under a plurality of predetermined conditions;
a second memory for storing, as a verification reference pattern corresponding to a registered speaker, a pattern which is obtained by said registered speaker speaking said password;
means for selecting a predetermined number of the control reference patterns from said first memory consecutively, starting from said control reference pattern which is the most similar to the verification reference pattern corresponding to said registered speaker;
means for determining, during verification, a first dissimilarity between the input pattern obtained by the speech of a person being verified and the verification reference pattern corresponding to said registered speaker, and second dissimilarities between said input pattern and the selected control reference patterns; and means for outputting a confirmation signal when said first dissimilarity is less than both said second dissimilarities and a predetermined threshold value.
2. A speaker verification system according to Claim 1, wherein the patterns which are obtained by speaking under the predetermined plurality of conditions are those which are obtained as a result of a plurality of predetermined persons speaking said password.
3. A speaker verification system comprising:
an input analyzer for analyzing a word spoken by a registered speaker or a person being verified, and outputting the result as an input pattern;
a registration number designator for outputting a registration number;
an input pattern buffer for storing said input pattern;
a verification reference pattern memory for storing said input pattern as a verification reference pattern during registration, designated by said registration number;
a control reference pattern memory for storing, as control reference patterns, patterns which are obtained when a plurality of predetermined persons speak a predetermined password;
a pattern matching unit for determining first dissimilarities between said verification reference pattern and said control reference patterns, a second dissimilarity between the input pattern of said person being verified, which is obtained during the verification, and said verification reference pattern, and third dissimilarities between said input pattern and said control reference patterns;
a control reference pattern designator for selecting a plurality of first dissimilarities consecutively, starting from the smallest of said first dissimilarities, to designate the control reference patterns corresponding to the selected dissimilarities as the control reference patterns corresponding to said registration number; and a judgement unit for generating a confirmation signal when said second dissimilarity is less than both said third dissimilarities corresponding to said designated control reference patterns and a predetermined threshold value.
an input analyzer for analyzing a word spoken by a registered speaker or a person being verified, and outputting the result as an input pattern;
a registration number designator for outputting a registration number;
an input pattern buffer for storing said input pattern;
a verification reference pattern memory for storing said input pattern as a verification reference pattern during registration, designated by said registration number;
a control reference pattern memory for storing, as control reference patterns, patterns which are obtained when a plurality of predetermined persons speak a predetermined password;
a pattern matching unit for determining first dissimilarities between said verification reference pattern and said control reference patterns, a second dissimilarity between the input pattern of said person being verified, which is obtained during the verification, and said verification reference pattern, and third dissimilarities between said input pattern and said control reference patterns;
a control reference pattern designator for selecting a plurality of first dissimilarities consecutively, starting from the smallest of said first dissimilarities, to designate the control reference patterns corresponding to the selected dissimilarities as the control reference patterns corresponding to said registration number; and a judgement unit for generating a confirmation signal when said second dissimilarity is less than both said third dissimilarities corresponding to said designated control reference patterns and a predetermined threshold value.
4. A speaker verification system according to Claim 3, further comprising a switch for connecting the output of said input pattern buffer to said verification reference pattern memory during registration and to said pattern matching unit during verification.
5. A speaker verification system comprising:
an input analyzer for analyzing first and second passwords spoken by a registered speaker or a person being verified, and outputting the result as input patterns;
a registration number designator for outputting a registration number;
an input pattern buffer for storing said input patterns;
a verification reference pattern memory for storing said input patterns as first and second verification reference patterns corresponding to said first and second passwords during registration, designated by said registration number;
a first control reference pattern memory for storing, as first control reference patterns, patterns which are obtained when a plurality of predetermined persons speak said first password;
a second control reference pattern memory for storing, as second control reference patterns, patterns which are obtained when said registered speaker speaks a plurality of words similar in pronunciation to said second password;
a pattern matching unit for determined first dissimilarities between said first verification reference pattern and said first control reference patterns, a second dissimilarity between a first verification input pattern corresponding to said first password obtained during verification and said first verification reference pattern, third dissimilarities between said first verification input pattern and said first control reference patterns, a fourth dissimilarity between a second verification input pattern corresponding to said second password and said second verification reference pattern, and fifth dissimilarities between said second verification input pattern and second control reference patterns;
a control reference pattern designator for selecting a plurality of first dissimilarities consecutively, starting from the smallest of said first dissimilarities, to designate the first control reference patterns corresponding to the selected dissimilarities as the first control reference patterns corredsponding to said registered number, and outputting said first control reference patterns to said pattern matching unit; and a judgement unit for generating a first confirmation signal when said second dissimilarity is less than both said third dissimilarities corresponding to said designated first control reference patterns and a predetermined threshold value, generating a second confirmation signal when said fourth dissimilarity is less than both said fifth dissimilarities and a predetermined threshold value, and generating another confirmation signal when both said first and second confirmation signals are generated.
an input analyzer for analyzing first and second passwords spoken by a registered speaker or a person being verified, and outputting the result as input patterns;
a registration number designator for outputting a registration number;
an input pattern buffer for storing said input patterns;
a verification reference pattern memory for storing said input patterns as first and second verification reference patterns corresponding to said first and second passwords during registration, designated by said registration number;
a first control reference pattern memory for storing, as first control reference patterns, patterns which are obtained when a plurality of predetermined persons speak said first password;
a second control reference pattern memory for storing, as second control reference patterns, patterns which are obtained when said registered speaker speaks a plurality of words similar in pronunciation to said second password;
a pattern matching unit for determined first dissimilarities between said first verification reference pattern and said first control reference patterns, a second dissimilarity between a first verification input pattern corresponding to said first password obtained during verification and said first verification reference pattern, third dissimilarities between said first verification input pattern and said first control reference patterns, a fourth dissimilarity between a second verification input pattern corresponding to said second password and said second verification reference pattern, and fifth dissimilarities between said second verification input pattern and second control reference patterns;
a control reference pattern designator for selecting a plurality of first dissimilarities consecutively, starting from the smallest of said first dissimilarities, to designate the first control reference patterns corresponding to the selected dissimilarities as the first control reference patterns corredsponding to said registered number, and outputting said first control reference patterns to said pattern matching unit; and a judgement unit for generating a first confirmation signal when said second dissimilarity is less than both said third dissimilarities corresponding to said designated first control reference patterns and a predetermined threshold value, generating a second confirmation signal when said fourth dissimilarity is less than both said fifth dissimilarities and a predetermined threshold value, and generating another confirmation signal when both said first and second confirmation signals are generated.
6. A speaker verification system according to Claim 5, further comprising a switch for connecting the output from said input pattern buffer to said verification reference pattern memory during registration, and to said pattern matching unit and said control reference pattern memory during verification.
7. A speaker verification system according to Claim 5, wherein said first and second passwords are different from each other, and wherein said second password is determined individually for each registered speaker.
8. A speaker verification system according to Claim 5, wherein said first and second passwords are same passwords.
Applications Claiming Priority (2)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| JP54132/1983 | 1983-03-30 | ||
| JP58054132A JPS59178587A (en) | 1983-03-30 | 1983-03-30 | Speaker confirming system |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CA1207907A true CA1207907A (en) | 1986-07-15 |
Family
ID=12962053
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CA000450839A Expired CA1207907A (en) | 1983-03-30 | 1984-03-29 | Speaker verification system |
Country Status (5)
| Country | Link |
|---|---|
| US (1) | US4694493A (en) |
| EP (1) | EP0121248B1 (en) |
| JP (1) | JPS59178587A (en) |
| CA (1) | CA1207907A (en) |
| DE (1) | DE3473957D1 (en) |
Families Citing this family (24)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| JPS63500126A (en) * | 1985-07-01 | 1988-01-14 | エッコ・インダストリ−ズ・インコ−ポレ−テッド | speaker verification device |
| US4910782A (en) * | 1986-05-23 | 1990-03-20 | Nec Corporation | Speaker verification system |
| DE3819178A1 (en) * | 1987-06-04 | 1988-12-22 | Ricoh Kk | Speech recognition method and device |
| US5144672A (en) * | 1989-10-05 | 1992-09-01 | Ricoh Company, Ltd. | Speech recognition apparatus including speaker-independent dictionary and speaker-dependent |
| US5127043A (en) * | 1990-05-15 | 1992-06-30 | Vcs Industries, Inc. | Simultaneous speaker-independent voice recognition and verification over a telephone network |
| US5365574A (en) * | 1990-05-15 | 1994-11-15 | Vcs Industries, Inc. | Telephone network voice recognition and verification using selectively-adjustable signal thresholds |
| US5125022A (en) * | 1990-05-15 | 1992-06-23 | Vcs Industries, Inc. | Method for recognizing alphanumeric strings spoken over a telephone network |
| US5303299A (en) * | 1990-05-15 | 1994-04-12 | Vcs Industries, Inc. | Method for continuous recognition of alphanumeric strings spoken over a telephone network |
| US5517558A (en) * | 1990-05-15 | 1996-05-14 | Voice Control Systems, Inc. | Voice-controlled account access over a telephone network |
| GB9021489D0 (en) * | 1990-10-03 | 1990-11-14 | Ensigma Ltd | Methods and apparatus for verifying the originator of a sequence of operations |
| US5271088A (en) * | 1991-05-13 | 1993-12-14 | Itt Corporation | Automated sorting of voice messages through speaker spotting |
| US5339385A (en) * | 1992-07-22 | 1994-08-16 | Itt Corporation | Speaker verifier using nearest-neighbor distance measure |
| CA2105034C (en) * | 1992-10-09 | 1997-12-30 | Biing-Hwang Juang | Speaker verification with cohort normalized scoring |
| US5430827A (en) * | 1993-04-23 | 1995-07-04 | At&T Corp. | Password verification system |
| WO1995005656A1 (en) * | 1993-08-12 | 1995-02-23 | The University Of Queensland | A speaker verification system |
| US5414755A (en) * | 1994-08-10 | 1995-05-09 | Itt Corporation | System and method for passive voice verification in a telephone network |
| DE19630109A1 (en) * | 1996-07-25 | 1998-01-29 | Siemens Ag | Method for speaker verification using at least one speech signal spoken by a speaker, by a computer |
| GB2316790B (en) * | 1996-08-30 | 2000-07-19 | Fujitsu Ltd | Principal identifying system, electronic settlement system, and recording medium to be used therefor |
| US7630895B2 (en) * | 2000-01-21 | 2009-12-08 | At&T Intellectual Property I, L.P. | Speaker verification method |
| US6012027A (en) * | 1997-05-27 | 2000-01-04 | Ameritech Corporation | Criteria for usable repetitions of an utterance during speech reference enrollment |
| US6510414B1 (en) | 1999-10-05 | 2003-01-21 | Cisco Technology, Inc. | Speech recognition assisted data entry system and method |
| GB2372366A (en) * | 2001-02-16 | 2002-08-21 | Imagination Tech Ltd | Speaker verification |
| JP4305509B2 (en) * | 2006-12-26 | 2009-07-29 | ヤマハ株式会社 | Voice processing apparatus and program |
| US9489950B2 (en) * | 2012-05-31 | 2016-11-08 | Agency For Science, Technology And Research | Method and system for dual scoring for text-dependent speaker verification |
Family Cites Families (10)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US31188A (en) * | 1861-01-22 | Drawer-alarm | ||
| US4053710A (en) * | 1976-03-01 | 1977-10-11 | Ncr Corporation | Automatic speaker verification systems employing moment invariants |
| JPS5569880A (en) * | 1978-11-22 | 1980-05-26 | Nec Corp | Pattern recognition unit |
| JPS56121100A (en) * | 1980-02-29 | 1981-09-22 | Nissan Motor | Voice identification driving device |
| AU7529981A (en) * | 1980-09-19 | 1982-03-25 | Hitachi Limited | Language analysis by pattern recognition |
| US4363102A (en) * | 1981-03-27 | 1982-12-07 | Bell Telephone Laboratories, Incorporated | Speaker identification system using word recognition templates |
| JPS5823097A (en) * | 1981-08-03 | 1983-02-10 | 日本電信電話株式会社 | Voice recognition apparatus |
| JPS5830800A (en) * | 1981-08-17 | 1983-02-23 | トヨタ自動車株式会社 | Voice recognition system for vehicle |
| JPS5876893A (en) * | 1981-10-30 | 1983-05-10 | 日本電気株式会社 | Voice recognition equipment |
| JPS6419597A (en) * | 1987-07-13 | 1989-01-23 | Toshiba Corp | Sample/hold circuit |
-
1983
- 1983-03-30 JP JP58054132A patent/JPS59178587A/en active Pending
-
1984
- 1984-03-20 US US06/591,652 patent/US4694493A/en not_active Expired - Lifetime
- 1984-03-29 CA CA000450839A patent/CA1207907A/en not_active Expired
- 1984-03-29 DE DE8484103502T patent/DE3473957D1/en not_active Expired
- 1984-03-29 EP EP84103502A patent/EP0121248B1/en not_active Expired
Also Published As
| Publication number | Publication date |
|---|---|
| US4694493A (en) | 1987-09-15 |
| EP0121248B1 (en) | 1988-09-07 |
| EP0121248A1 (en) | 1984-10-10 |
| JPS59178587A (en) | 1984-10-09 |
| DE3473957D1 (en) | 1988-10-13 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CA1207907A (en) | Speaker verification system | |
| US5895447A (en) | Speech recognition using thresholded speaker class model selection or model adaptation | |
| EP1019904B1 (en) | Model enrollment method for speech or speaker recognition | |
| US4581755A (en) | Voice recognition system | |
| US5339385A (en) | Speaker verifier using nearest-neighbor distance measure | |
| KR100297833B1 (en) | Speaker verification system using continuous digits with flexible figures and method thereof | |
| US8433569B2 (en) | Method of accessing a dial-up service | |
| KR19980086697A (en) | Speaker recognition method and apparatus in speech recognition system | |
| JPS6217240B2 (en) | ||
| US6556969B1 (en) | Low complexity speaker verification using simplified hidden markov models with universal cohort models and automatic score thresholding | |
| US20080071538A1 (en) | Speaker verification method | |
| WO1986005618A1 (en) | Individual recognition by voice analysis | |
| Marinov | Text dependent and text independent speaker verification system: Technology and application | |
| US9978373B2 (en) | Method of accessing a dial-up service | |
| JP2000148187A (en) | Speaker recognizing method, device using the method and program recording medium therefor | |
| KR19980045013A (en) | How to Improve Speaker Recognizer by Entering Password | |
| JPH02149900A (en) | Voice recognizing and answering device | |
| KR100349656B1 (en) | Apparatus and method for speech detection using multiple sub-detection system | |
| JPH01302297A (en) | Speaker recognition device | |
| JPH0344316B2 (en) | ||
| JPH0431896A (en) | Speech recognizing device | |
| JPS61249098A (en) | Speaker collation system | |
| Xu et al. | Rejection in speech recognition based on CDCPMs | |
| JPH01185600A (en) | Speech recognizing device | |
| JPS62249199A (en) | Speaker checking system |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| MKEX | Expiry |