GB2403327A - Identity verification system with speaker recognition - Google Patents

Abstract

A method of speaker recognition comprising receiving a user identifier, receiving a user utterance, if user identifier has speaker adapted speech recognition model associated with it then use this model to recognise user utterance, else use a speaker independent speech recognition model to recognise user utterance, and make a speaker adapted model associated with the user identifier, then, during use, the speaker adapted speech recognition model is further adapted to the speaker. In particular the user utterance is received after the system interrogates the user with a question formulated to elicit a response within a specific range of possible numerical responses.

Description

1 2403327

METHOD OF ADAPTING A SPEECH-BASED IDENTITY VERIFICATION SYSTEM

Field of the Invention

The invention relates to Identity venfcaton systems using speech as the basis of venfyng the Identity of a user

Background

Identity venfcaton systems using speech from a user to verify the identify of that user are conventionally based on either speech recognition or speaker recognition methodologies Identity verification systems using speech from a user to verify the Identify of a user are used as a means of restricting access to a secure system or area, to authonsed users If, for example, a user Is determined to be an authonsed user, the occurrence of some subsequent operation such as enabling physical or remote access to a user may be permitted Common applications Include door access control, bank and phone accounts access control and remote network access control Such systems are constantly subjected to attack by unauthorsed users and means of Improvement of their security are constantly sought Speech recognition based identity verification systems using speech from a user to verify the identify of a user seek to determine the words that the user has spoken as the basis of verifying the Identity of the user Such systems are capable of recognsng speech comprising words and numbers and may be speaker dependent or speaker Independent Speaker recognition based Identity venfcaton systems using speech from a user to verify the Identify of a user, rely on previously saved Information that represent particular vocal characteristics of the user seeking dentfcaton More specifically, the speech signal which results from a users speech Is analysed to extract certain distinctive features of that speech signal These features are In turn compared with corresponding features that have been extracted from previously recorded speech samples spoken by the same user The user's identity Is venfed If the results of such comparisons meet predetermined criteria Such schemes are capable of making very accurate comparisons As speaker recognition based security schemes are able to offer comparatively better levels of security than speech recognition identification schemes it Is desirable for providers of secure services to adopt such schemes However there are disadvantages of adopting speaker recognition based Identity verification schemes Disadvantages Include the requirement for tme-consumng speaker- specfc enrollment sessions before an authonsed user can successfully use such systems Training sessions Involve the user providing voice samples that are used to create voice models, which are compared against recorded speech signals provided at the time a user seeks Identity venfcaton. A further disadvantage Is the susceptbity of such systems to 'replay- a/tack' A replay-attack Involves an unauthorsed third party using a recording of an authorsed user's voice to fool an dentfcaton system Into concluding that an authorsed usems attempting to access the secure system Another disadvantage Is that organsatons providing services secured by Identity venfcaton systems are disinclined to use systems that place any additional burden upon the user beyond that required by speech recognition based Identity venfcaton schemes The present Invention seeks to overcome some the problems described above and provide a method of adapting an Identity venfcaton scheme relying on speech recognition methodologies, over time, to a scheme whereby Identity venfcaton Is based upon speaker recognition

Summary of the Invention

According to the present Invention there is provided a method of adapting an Identity verification system based on speech recognition, over time, to an Identity verification system based on speaker recognition, comprising the steps of receiving a user Identifier from a user seeking identity verification, formulating and making a challenge to the user based on information available to the Identity verification system associated with said usemdentfer, receiving a speech signal from the user, determining whether a voice model associated with said user exists corresponding to the user response to said challenge, and If said voice model does exist compare voice model with speech signal to verify the Identity of the user, and If said voice model does not exist assess said speech signal to determine If Information contained In said speech signal contains Information associated with the user to verify the Identity of the user, and If the user's Identity Is venfed on the basis of said assessment said speech signal Is used In the creation of a voice model for use for future Identity venfcation of the user Advantages of this method Include the fact that an identity verification system will, over time, acquire speaker-recognton functionality thereby enhancing the security of the identity verification system, tme- consumng speaker specific enrollment sessions are not required to take place before an authonsed person's Identity Is able to be verified by the system on a speaker recognition basis, no noticeable additional burden beyond that required by a speech based Identity venfcaton scheme Is placed upon users, and once Implemented, substantially reduces the number of challenges a user seeking access would face, thus reducing the load on the Identity verification system as a whole (and on the user), which would In turn enable the Identity venfcaton system to verify the identity of more users In the same period of time than that of a conventional speech recognition based Identity verification system According to the present invention there Is also provided a method of adapting an Identity venfcaton system based on speech recognition, over time, to an Identity verification system based on speaker recognition, comprising the steps of receiving a user Identifier from a user seeking identity verification, formulating a challenge to the user based on information associated with said user identifier to elicit a vocal response from the user within a specific range of possible responses, making said challenge to the user, receiving a speech signal from said user, determining whether a voice model associated with the user exists corresponding to the user response to said challenge, and If said voice model does exist compare voice model with speech signal to verify the identity of the user, and If said voice model does not exist assess said speech signal to determine If Information contained In said speech signal contains Information associated with the user to verify the Identity of the user, and If the user's Identity Is venfed on the basis of said assessment said speech signal Is used In the creation of a voice model for use for future Identity verification of the user Limiting the range of available user responses Is particularly desirable for speaker recognition schemes The advantage being that only a limited number of user-specfc voice models would need to be created and stored In order to carry out comparisons of a speech sample acquired from a user at the time of Identity venficaton, with the stored voice models According to the present Invention there Is further provided a method of adapting an Identity verification system based on speech recognition, over time, to an Identity venfcaton system based on speaker recognition, comprising the steps of receiving a user Identifier from a user seeking Identity venfcaton, formulating a challenge to the user based on information associated with said user Identifier to elicit a vocal response from the user within a specific range of possible numerical responses, making said challenge to the user, receiving a speech signal from said user representing a number within a specific numerical range, determining whether a voice model associated with the user exists corresponding to the user response to said challenge, and if said voice model does exist compare voice model with speech signal to verify the Identity of the user, and if said voice model does not exist assess said speech signal to determine If Information contained In said speech signal contains Information associated with the user to verify the identity of the user, and If the user's Identity Is venfed on the basis of said assessment said speech signal Is used In the creation of a voice model for use for future identity venfcaton of the user.

Limiting the responses of users to a specific range of numbers not only reduces the number of user-specfc voice models required In order to carry out comparisons of a speech sample acquired from a user at the time of Identity venfcaton, with the stored voice models, but also offers the opportunity to reduce the identity venfcaton system vulnerability to a 'replay-attack' Limiting a user's responses to numbers corresponding to specific aspects of a user's personal Information (as opposed to responses directly corresponding to a user's personal Information or requiring a user to vocalse personal information) to verify the identity of that user reduces the probability of a third party recording a user's responses and using the same to fool an Identity venfcaton system, or gain personal information concerning that user

Description of the Invention

An embodiment of the present Invention will now be described by way of example only, with reference to the accompanying drawings In which figure 1 represents an access control system using a method of performing Identity verification figure 2 represents an access control system using speech acquired from a user to verify the Identity of a user figure 3 represents an access control system using speech acquired from a user to verify the Identity of that user figure 4 represents an Identity analysis sub-system 5 shown In figure 2 based solely on speech recognition to verify the Identity of a user figure 5 represents a method of adapting an identity verification system based on speech recognition, over time, to an Identity venfcaton system based on speaker recognition Referring to figure 1, an access control system for performing a method of performing Identity venfcaton Is shown The access control system limits access by a user 1 to a secure system 3 by an Identity verification system 2 The secure system 3 may be a remote network Refemng to figure 2, an access control system using speech acquired from a user 1 to verify the Identity of that user Is shown with a speech signal acquisition subsystem 4 and an Identity analysis subsystem 5 Typically a user seeking access to a secure system 3, at the time that that user seeks access to the secure system 3 Will provide a unique user Identifier, such as for example, a personal dentfcaton number Subsequently, a user's vocal responses to challenges are acquired using a speech acquisition subsystem 4 Those vocal responses are processed by an Identity verification subsystem 5 to determine that the unique user identifier corresponds to the user 1 seeking access, prior to making an identity venfcaton decision and permitting or denying access by the user 1 to the secure system 3 Refernng to figure 3, an access control system with a user 1 and an identity verification system 2 comprised of a speech acquisition subsystem 4 and an identity analysis subsystem 5 and a secure system 3 are shown A user 1 submits a unique user identifier that Is received 6 by a speech acquisition subsystem 4 If the unique usemdentfer Is known to the speech acquisition subsystem 4, information associated with that user Is accessed and an appropriate challenge formulated 7 If the unique user identifier Is unknown to the speech acquisition subsystem 4, the Identity verification request may be denied and the process ends It will be appreciated that the speech acquston subsystem 4 may In an alternative embodiment within the scope of the present Invention, require the user to re-enter the unique use Identifier and so on, until the Identity verification system receives a valid Identifier or the user's attempts to do so reach a maximum permitted number, whereby for example, the identity verification process may be denied Upon receipt of a valid unique user Identifier, the speech acquisition subsystem will formulate a challenge 7 based on Information known to both the user 1 and the speech acquisition subsystem 4, corresponding to the unique usemdentfier Preferably the challenge 7 Is randomly formulated Preferably, the challenge Is formulated to elicit a vocal response falling within a specific range of possible responses As explained below In further detail, this Is particularly desirable for speaker recognition schemes as only a limited number of user-specfc voice models need to be created and stored In order to carry out comparisons of a speech sample acquired from a user at the time of Identity venfcaton, with the stored voice models Preferably, the challenge is formulated to elicit a vocal response limited to a specific range of numbers (for example numbers 0 to 9) that correspond to specific aspects of a user's personal Information (as opposed to responses directly corresponding to a user's personal information) By way of example only, If the information used to formulate the challenge Is a password ROBERT, such a challenge could be - "At what character number does the fetter 'B' occur In your password?" Where speech recognition is used within an identity verification system it Is desirable that the maximum range of responses are available to the user within a minimal vocabulary Another advantage of using such an approach Is that there would not be any need to retrain a speech recognition model should a user adopt a new unique personal identifier It will appreciated that a great number of possible challenges can be formulated based on nformabon the system associates with that unique usemdentfier The speech acquisition subsystem Issues the challenge that has been formulated 8 to the user 1 Preferably the challenge Is made audibly, possibly via a portable digital device such as a cellular phone connected to a telecommunications network It Will appreciated that methods of making the challenge could be made other than audibly For example, visually, If the means were available to present the challenge to the user 1 In such a visual manner It will be appreciated that a number of challenges could be made to a user 1 Multiple challenges could for example, enhance the level of security desired by the identity verification system 2 In response to the challenge 8 the user 1 responds vocally 9 Preferably the response received Is a number By way of example only, If the Information used to formulate the challenge Is a password ROBERT, and a challenge made to the user Is "At what character number does the letter 'B' occur In your password?", an appropriate response would be for the user to utter the word "3" The user's vocal response 9 Is recorded 10 by the speech acquisition subsystem 4 It will be appreciated that a number of speech responses could be acquired from a user For example, if multiple challenges were to be put by the speech acquisition subsystem 4 to the user 1 Referring to figure 4, an Identity analysis subsystem 5 shown In figure 2 based on speech recognition means Is shown A recorded speech signal representing a user's vocal response 10 Is analysed by a word recognition process 11 to determine If words can be identified within the speech signal Preferably the word recognition processing 11 is limited to recognsng a specific range of words Word recognition techniques are capable of recognsng speech comprising words or numbers In order to achieve accurate recognition of words, due to the fact that electronic speech recognition means have problems dstngushng between phonetically similar speech, lengthy sessions Involving a user may be necessary to train such systems to make accurate distinctions Identification of numbers on the other hand, are comparatvely'easer' for such systems Therefore In one particularly desirable embodiment, it Is preferable that verbal responses be limited to a set of numbers For example the numbers 0 to 9 If the word recognition analysis 11 falls to recognse any words within the recorded speech signal 10, the user's Identity Is not validated 14 and the user 1 Is denied access 15 to the secure system 3 It will be appreciated that In an alternative embodiment, that If the word recognition analysis 11 falls to recognse any words within the recorded speech signal 10, rather than denying the user access to the secure system 15, further challenges could be put to the user to obtain further user speech signals If the analysis 11 recognses words, further analysis 12 Is undertaken to determine If those words correspond to the challenge 8 If the further analysis 12 determines that the words recognsed at the word recognition analysis 11 do not correspond to the challenge 8, the user's Identity Is not validated 14 and the user 1 Is denied access to the secure system 3 It Will be appreciated that In an alternative embodiment, that If the further analysis 12 determines that the words recomposed at the word recognition analysis 11 do not correspond to the challenge 8, rather than denying the user access to the secure system 15, further challenges could be put to the user to obtain further user speech signals until the speech analysis subsystem receives a speech signal 10 containing words corresponding to the challenge 8 or until the user's attempts to do so reach a maximum permitted number, whereby for example, the user 1 could be denied access to the secure system 3 If the further analysis 12 determines that the words recognised 11 correspond to the challenge 8 the user's Identity Is validated 13 and the user Is permitted access 16 to the secure system 3 It will be appreciated that In alternative embodiments requiring higher security levels, If the words recognsed 11 correspond to the challenge 8, rather that validating the user's Identity 13, a multiplicity of speech signals and could be obtained by the speech acquisition subsystem 4 and a multiplicity of speech signal analyses could be undertaken by the Identity analysis subsystem 5 Refemng to figure 5, a method of migrating an Identity analysis subsystem based on speech recognition (as shown In figure 4) to speech recognition means over time, Is shown A recorded speech signal representing a user's vocal response 10 Is analysed 19 to determine If a voice model associated with the user for the user's response exists If a voice model associated with the user for the user's response 19 does not exist the recorded speech signal 10 Is analysed by a word recognition process 11 to determine If words can be Identified within that speech signal 10 If the analysis 11 falls to recognse any words within the recorded speech signal 10, the user's Identity is not validated 14 and the user 1 Is denied access 15 to the secure system 3 If the analysis 11 recognses words, further analysis 12 Is undertaken to determine If those words correspond to the challenge 8 If the further analysis 12 determines that the words recognsed at the word recognition analysis 11 do not correspond to the challenge 8, the user's identity Is not validated 14 and the user 1 Is denied access to the secure system 3 If the further analysis 12 determines that the words recognsed 11 correspond to the challenge 8 the user's Identity Is validated 13 If a user's identity Is validated 13 a voice model 17 Is created based on the recorded speech signal 10 and the user Is permitted access to the secure system 16 and the voice model 17 Is stored In the voice model database 18. If the further analysis 12 determines that the words recognsed at the word recognition analysis 11 do not correspond to the challenge 8, the user's Identity Is not validated 14 and the user 1 is denied access 15 to the secure system 3 If a voice model associated with the user for the user's response 19 exists, that voice model Is retrieved from the voice model database 18 and compared 20 with the recorded speech signal 10 If the comparison 20 Is unsuccessful the user's Identity Is not validated 14 and the user 1 Is denied access to the secure system 3 The purpose of updating existing voice models Is to track changes In a users voice charactenstcs over time If the comparison 20 Is successful, the voice model associated with the user for the user's response 19 Is updated 17, the updated voice model 17 Is stored In the voice model database 18 and the user Is permitted access 15 to the secure system 13.

Preferably the Identity analysis subsystem operations are limited to creating and storing voice models for a specific range of words, for example the numbers O to 9 Each time a user's Identity Is verified by the system described herein, further voice models are created or updated, eventually building a complete library of voice models for future speaker recognition based Identity verification It Will be appreciated that In an alternative embodiment requiring higher security levels, that If the words recognsed 11 correspond to the challenge 8, rather that validating the user's identity 13, a multiplicity of speech signals and could be obtained by the speech acquisition subsystem 4 and a muitplcty of speech signal analyses could be undertaken by the Identity analysis subsystem 5 It will be appreciated that If the Identity acquisition subsystem 4 operates to eict a limited number of specific responses from the user 1 (for example the words 0 to 9), over time, this scheme (as represented In figure 5) offers a means of adapting an Identity verification scheme from relying on a speech recognition methodology, to a scheme whereby ultimately the Identity verification Is wholly based upon speaker recognition means Likewise it will be appreciated that such a scheme would not from a users perspective, be apparent, it would not place any additional burden upon such users In the creation of such enhanced security capabilities, and once Implemented, would offer higher levels of security while substantially reducing the number of challenges a user seeking access would face, thus reducing the load on the Identity verification system as a whole (and on the user), which In turn would allow an Identity venficaton system to verify the Identity of more users In the same period of time than that of an Identity venfcaton system relying on speech recognition methodologies

Claims (1)

1. Claims 1 A method of adapting an Identity verification system based on

   speech recognition, over time, to an Identity verification system based on speaker recognition, comprising the steps of receiving a user Identifier from a user seeking Identity verification, formulating and making a challenge to the user based on Information available to the Identity venfcaton system associated with said usemdentfer, receiving a speech signal - from the user, determining whether a voice model associated with said user exists corresponding to the user response to said challenge, and If said voice model does exist compare voice model with speech signal to verify the identity of the user, and If said voice model does not exist assess said speech signal to determine if information contained In said speech signal contains information associated with the user to verify the identity of the user, and If the user's identity Is verified on the basis of said assessment said speech signal Is used In the creation of a voice model for use for future Identity verification of the user 2 A method of adapting an Identity verification system based on speech recognition, over time, to an Identity venfcaton system based on speaker recognition, comprising the steps of receiving a user Identifier from a user seeking identity verification, formulating a challenge to the user based on information associated with said user identifier to elicit a vocal response from the user within a specific range of possible responses, making said challenge to the user, receiving a speech signal from said user, determining whether a voice model associated with the user exists corresponding to the user response to said challenge, and If said voice model does exist compare voice model with speech signal to verify the Identity of the user, and If said voice model does not exist assess said speech signal to determine If information contained In said speech signal contains Information associated with the user to verify the Identity of the user, and If the user's identity Is venfed on the basis of said assessment said speech signal Is used In the creation of a voice model for use for future identity verification of the user 3 A method of adapting an identity verification system based on speech recognition, over time, to an Identity venfcaton system based on speaker recognition, comprising the steps of receiving a user Identifier from a user seeking Identity verification, formulating a challenge to the user based on information associated with said user identifier to elicit a vocal response from the user within a specific range of possible numerical responses, making said challenge to the user, receiving a speech signal from said user representing a number within a specific numerical range, determining whether a voice model associated with the user exists corresponding to the user response to said challenge, and If said voice model does exist compare voice model with speech signal to verify the identity of the user, and If said voice model does not exist assess said speech signal to determine if Information contained In said speech signal contains information associated with the user to verify the Identity of the user, and if the user's identity Is venfed on the basis of said assessment said speech signal Is used In the creation of a voice model for use for future Identity verification of the user 4 IN computer program composing program instructions for causing a computer to adapt an identity verification system based on speech recognition, over time, to an identity verification system based on speaker recognition according to any preceding claim 5. Apparatus for performing a method according to any preceding claim