Europaisches Patentamt 19 European Patent Office

Office europeen des brevets © Publication number: 0 654 781 A2

12 EUROPEAN PATENT APPLICATION

@ Application number: 94308231.3 @ Int. CI.6: G10L 5/06

(22) Date of filing : 09.11.94

(So) Priority: 19.11.93 US 155973 @ Inventor: Sachs, Richard M. 64 Sunset Place @ Date of publication of application : Middletown, New Jersey 07748 (US) 24.05.95 Bulletin 95/21 Inventor : Schoeffler, Max S. 17 Kenwood Lane @ Designated Contracting States : Matawan, New Jersey 07747 (US) DE ES FR GB IT (74) Representative : Watts, Christopher Malcolm @ Applicant : AT & T Corp. Kelway, Dr. et al 32 Avenue of the Americas AT&T (UK) Ltd. New York, NY 10013-2412 (US) 5, Mornington Road Woodford Green Essex, IG8 0TU (GB)

(S) Method of accommodating for carbon/electret set variability in automatic speaker verification.

In verification method of (57) a speaker system, a FIG. 6 compensating for differences in speech sam- ples obtained during registration and those obtained during verification due to the use of VERIFICATION (4-WAY) different types of is provided by ,601 at least of the such that filtering one samples PROMPT the similarities of the two samples are in- creased. The filtered sample is used within the ,604 speaker verification matching process. A two- RECEIVE VERIFICATION way comparison is disclosed in which both a SPEECH SAMPLE verification speech sample and a reference sample are filtered with nonlinear r609 ,606 characteristics such as carbon microphone PRODUCE CARBON characteristics. A is also VERIFICATION FILTER four-way comparison PATTERN SAMPLE disclosed in which patterns produced from un- filtered verification and reference samples and 611 from the filtered verification PRODUCE CARBON patterns produced FILTERED and reference samples are compared to identify VERIFICATION a match. A score is determined for each com- PATTERN The the best DB (407) parison. comparison having score 613 is used to determine if a match has occurred. 6-i 4-WAY COMPARISON WITH TWO STORED REFERENCE PATTERNS 617 DB (412) PICK "BEST" SCORE CM r619 r622 < CLOSE ENOUGH?^ ^ SECOND TRY ' I YES I YES 00 \ /-624,624„. I --627 GRANT ACCESS DENY 10 /633 CO ( BEGIN SERVICE") C END CALL) 635 C END ) LU Jouve, 18, rue Saint-Denis, 75001 PARIS 1 EP 0 654 781 A2 2

Background of the Invention tion pattern generated from a speech sample provid- ed by an imposter. The present invention is generally directed to Thus, a subscriber who registers using one type speaker verification, and more particularly, to a meth- of telephone handset microphone and attempts to be od of accommodating variability among different 5 "verified" using another type of handset microphone types of telephone handsets, in order to improve the is more likely to be denied access than one who reg- accuracy of speaker verification. isters and attempts to be verified using the same type Speaker Verification (SV) is a speaker-depend- of handset microphone. ent pattern-matching process in which a subscriber's speech sample presented for verification is process- 10 Summary of the Invention ed to produce a verification pattern. This verification pattern is compared to an SV reference pattern that In accordance with the present invention, the is typically produced from speech samples previously problem of compensating for variability in speech provided in the course of a so-called registration ses- samples due to the use of different types of micro- sion. A "match" between the verification and refer- is phones is solved by filtering at least one of the sam- ence patterns occurs when their characteristics are ples in accordance with the characteristics of one of substantially similar. Otherwise, a "mismatch" is said the microphone types and using the filtered sample to have occurred. within the matching process. A typical application of SV is a -based In general, it is not possible to determine whether security system. Asubscriber "registers" with the sys- 20 any particular speech sample originated from any tem by providing speech samples over a telephone particular type of microphone. Therefore, in preferred link and an SV reference pattern is produced. Subse- embodiments, both the verification speech sample quently, a caller, seeking access to, for example, a and the SV reference sample are filtered with typical service or some secure data, calls the system and carbon-microphone characteristics. Consequently, presents his/herspeech sample forverification as de- 25 any variability which may have resulted from using scribed above. If a match occurs, the desired access different types of handset microphones is reduced. is granted. If there is a mismatch, it is presumed that Variability originating from other properties of the a so-called imposter-pretending to be a subscriber- speech sample such as added background , and was the caller and access is denied. telephone network distortion or variable speaking lev- Many times, SV is complicated by the fact that the 30 el is also reduced. For example, if the samples are verification pattern is different from the SV reference generated by an electret microphone, the filtering pattern due to circumstances such as, illustratively, causes the samples to have similar characteristics to the use of different types of telephone handset micro- samples that would have been generated by a carbon phones, e.g., linear (such as electret) and non-linear microphone. If the samples are generated by a carbon (such as carbon). Other examples include different 35 microphone, the filtering will result in samples which, background noises and different speaking levels. although now different, retain their essential charac- These differences can cause characteristics of the ter as carbon microphone speech samples. Thus, no speech sample provided during registration and the matter which type of microphone was used to provide speech sample provided during any particular SV ver- the two samples, theirf iltered versions both have car- ification session to be different from one another. The 40 bon-microphone-like characteristics. corresponding patterns will then also be different, The principal consequence of the foregoing is possibly resulting in an incorrect "mismatch" determi- that because the invention reduces the variability be- nation. tween samples provided using different microphone In particular, an electret microphone performs a types, that variability need not be taken into account fairly linear transformation on incoming speech sam- 45 when establishing criteria under which a "match" will ples and, as such, minimally distorts them. A carbon occur. Indeed, the invention allows those criteria to be microphone, on the other hand, performs a non-linear made more stringent while not increasing the level of transformation on the speech samples by, for exam- incorrect rejection (the latter being the declaration of ple, compressing high-volume speech levels and sup- a mismatch when the caller is, in fact, the subscriber). pressing low background noise levels, the latter often so In an alternative embodiment of the invention, being referred to in the art as "enhancement." As patterns produced from unf iltered versions of the ver- such, the carbon microphone distorts the speech ification and reference samples are used along with samples to a significant extent. Because of the vari- the patterns produced from the filtered versions of ability in the effects that these different types of mi- the verification and reference samples as described crophones have on the samples, it is difficult to dis- 55 above. Comparisons are made between each version criminate between a mismatch caused by using dif- of the verification pattern and each version of the ref- ferent types of microphones and a mismatch caused erence pattern. The results are then used to deter- by comparing an SV reference pattern to a verifica- mine whether a match has occurred. This approach 2 3 EP 0 654 781 A2 4 could, in theory, improve the overall system perfor- form 10 which implements the principles of the pres- mance, for reasons that are explained in detail here- ent invention. At the heart of service platform 10 is a inbelow. microprocessor 11 and various standard peripherals Variability in the patterns can arise from factors with which it communicates over bus 13. The periph- other than differences in microphone type. For exam- 5 erals include random access memory (RAM) 12, ple, background noise derived acoustically or from read-only memory 14, hard disk memory 16, tele- telephone-network-based circuitry may introduce va- phone interface 18, digital signal processor (DSP) 19 riability into the patterns. Other factors such as vari- and a number of other peripherals indicated at 15. (Al- able speaking level or variability arising from other though not shown in the FIG., DSP 19 may have its properties of the utterance not related to speaker dif- 10 own memory elements and/or a direct connection to ferences may also introduce variability which may re- various memory elements within the system, such as sult in a mismatch determination. disk memory 16.) Indeed, the principles of the invention can be Service platform 10 is accessible only by sub- used to address such other variabilities. In particular, scribing individuals referred to herein as "subscrib- the invention generally encompasses the concept of 15 ers." The process of becoming a subscriber includes processing at least one of the recognition and verifi- a "registration" process wherein the subscriber is cation speech samples so that the properties charac- asked to recite utterances which are converted into terizing the processed speech sample are more sim- reference speech samples. This is illustratively car- ilar to the properties of the other speech sample than ried out during a telephone call made to the system is the unprocessed speech sample. The processing 20 from rotatory telephone set 31 via telephone central could thus be noise-reduction processing or volume- office (CO) 20 and a telephone line 21 extending from normalization processing, or whatever processing is CO 20 to telephone interface 18. Those reference available to make the recognition and verification speech samples are used to produce a speaker veri- patterns of the same subscriber are closer to one an- fication (SV) reference pattern for each registered other. 25 subscriber, those patterns being stored in a data sec- tion of disk memory 16, and indicated as database Brief Description of the Drawing 168. The routines which produce the reference pat- tern from the speech samples are pattern production FIG. 1 illustrates a block diagram of a computer routines 163. system service platform in accordance with the pres- 30 Assume, now, that at a later time the subscriber ent invention. is connected to service platform 10 from rotary tele- FIG. 2 illustrates three dimensional representa- phone 31, or from touch-tone telephone 32 also con- tions of a reference pattern and verification pattern nected to CO 20 or, indeed, from any other telephone generated by an electret microphone and a carbon connected to any other central office. Once the call microphone filtered with carbon-microphone charac- 35 has been answered by the service platform 10, a teristics. process of verifying the subscriber is initiated. That FIG. 3 illustrates a table representing the condi- process, which is embedded in SV routine 161, tions in which a closest match will occur as a function prompts the call for a subscriber identification num- of the reference pattern and verification pattern. ber, which may be entered either via touch-tone entry FIG. 4 illustrates a flow chart depicting a method 40 or speaker-independent digit recognition. This en- of registering a reference speech sample. ables the platform to retrieve the stored reference FIG. 5 illustrates a flow chart depicting a method pattern for the subscriber in question. The subscriber of verifying a verification speech sample using a one- is prompted to recite a desired utterance. As was the way comparison in accordance with the present in- case during registration, the recited utterance is con- vention. 45 verted to a set of speech samples from which a veri- FIG. 6 illustrates a flow chart depicting a method fication pattern is produced again using pattern pro- of verifying a verification speech sample using a four- duction routine 163. If a comparison of the two pat- way comparison in accordance with the present in- terns indicates that the caller is, indeed, the subscrib- vention. er, the caller is granted access to the service platform FIG. 7 illustrates a block diagram which depicts so 1 0. The service platform 1 0 may, for example, serve a carbon filter process in accordance with the present as a source of information to which only subscribers invention. have access. Voice mail is a typical example. Or, the FIG. 8 illustrates a memoryless transfer function platform may be a gateway to an outbound calling of the carbon filter of FIG. 7. service, or may be a resource used by another ser- 55 vice wanting to verify a person's identity. Detailed Description Many times, SVis complicated by the fact that the verification pattern is different from the SV reference FIG. 1 shows a computer system service plat- pattern due to circumstances such as, illustratively, 3 5 EP 0 654 781 A2 6 the use of different types of telephone handset micro- rejections. Use of the more stringent criteria, more- phones. In the present illustrative embodiment, for over, means an enhanced ability to reject imposters. example, rotary telephone 31 has a non-linear, car- In general, it is not possible to determine whether bon microphone while touch-tone telephone 32 has a any particular speech sample originated from any linear, electret microphone. These differences can 5 particular type of microphone. Therefore, in preferred cause characteristics of the speech sample provided embodiments, both the verification speech sample during registration and the speech sample provided and the SV reference sample are filtered. Conse- during any particular SV verification session to be dif- quently, any variability which may have resulted from ferent from one another. The corresponding patterns using different types of handset microphones is re- will then also be different, resulting in an incorrect 10 duced. For example, if the samples are generated by "mismatch" determination. an electret microphone, the filtering causes the sam- Variability in the patterns can arise from other ples to have similar characteristics to samples that factors as well. Background noise derived acoustical- would have been generated by a carbon microphone. ly or from telephone-network-based circuitry may in- This was just described in connection with patterns troduce variability into the patterns. Other factors 15 201 , 202 and 203. If the samples are generated by a such as variable speaking level or variability arising carbon microphone, the filtering will result in samples from other properties of the utterance unrelated to the which, although now different, retain their essential caller's identity may also introduce variability which character as carbon microphone speech samples. may result in a mismatch determination. This is illustrated by pattern 204, which is the pattern These problems can be more fully appreciated by 20 that results by filtering pattern 202 with the carbon- a consideration of FIG. 2. This FIG. shows represen- microphone characteristic. Thus no matter which type tations of a reference pattern 201 produced from an of microphone was used to provide the reference and electret-microphone-originated utterance spoken by verification samples, their filtered versions both have a particular subscriber and a verification pattern 202 carbon-microphone-like characteristics. Advanta- produced when that same subscriber spoke the same 25 geously, this directly enables the platform 1 0 to make utterance but from a telephone having a carbon mi- the criteria for matching more stringent, as noted crophone. As is well known in the art, and as depicted above. in the FIG., these patterns can be thought of as a time The overall process as just described is repre- sequence of feature vectors, only one of which is de- sented by the flowcharts of FIGS. 4 and 5. FIG. 4, in picted explicitly for each pattern. Note that the pat- 30 particular, represents the registration process. As in- terns are different, resulting from the different char- dicated at 401, the subscriber is prompted to recite acteristics of the two microphone types. These differ- the utterance. Reference speech samples are gener- ences in the patterns may result in an SV mismatch. ated from the received utterances (402). Those sam- In the prior art, to achieve a particular desired level of ples are carbon-filtered in accordance with the inven- verification acceptance, the criteria by which a match 35 tion (404), the reference pattern is produced (406), is determined must be sufficiently relaxed to compen- and stored (407) in a portion of database 168. As sate for this phenomenon. Disadvantageously, how- shown in FIG. 4, that portion of the database is refer- ever, such relaxation will make it easier for an impos- red to as DB(407). ter to gain access. FIG. 5 represents the verification process which, The present invention alleviates this problem. In 40 for reasons that will become apparent, is referred to particular, the variability in speech samples due to the as the one-comparison process. A verification pat- use of different types of microphones is dealt with by tern is produced in steps 501 , 502, 504 and 507 which filtering at least one of the samples in accordance parallel steps 401, 402, 404 and 406, respectively, in with the characteristic of one of the microphone types FIG. 4. (Although not explicitly shown, the prompting and using the filtered sample within the matching 45 step 501 includes prompting the caller for the above- process. However, the general methods described mentioned identification number.) In step 511, the below could also be used to reduce other types of va- verification pattern produced in step 507 is compared riability which may arise from certain properties of the to the reference pattern retrieved from DB(407) speech samples. A routine which carries out this type (based on the subscriber identification number). If, at of filtering is denoted in FIG. 1 at 164, with the data so step 514, the two patterns are "close enough" to one defining the filter characteristic being denoted at 167. another-based on well-known speech verification Thus as further shown in FIG. 2, the electret-or- methods-access to the service provided by the plat- iginated pattern 201, is filtered in accordance with a form is granted (521), service begins (522), and the carbon microphone characteristic, resulting in a car- process ends (524). bon-filtered pattern 203 which is closer in similarity to 55 If, on the other hand, the two patterns are not pattern 202 than is pattern 201. It is therefore possi- "close enough" to one another, and if this is not the ble to make the acceptance criteria more stringent caller's second attempt to gain access during this call, than before without increasing the level of incorrect as determined at 515, the process returns to step 4 7 EP 0 654 781 A2 8

501 , thereby giving the caller a "second try." If it is the to verify the subscriber. caller's second attempt, access is denied (517) and The parameters of the four-comparison case are the platform terminates the call (hangs up). It is pos- set forth in FIG. 3. Each line entry represents one of sible that the caller is, in fact, the subscriber but for the four combinations of patterns produced from fil- any of a number of reasons was denied access any- 5 tered and unfiltered reference samples, and filtered way. Accordingly, an alternative to terminating the and unfiltered verification samples. Depending on call is to transfer it to a live operator who can verify what type of microphone was actually used in any the caller's identity using information as might be particular case to produce the reference samples and available such as the caller's mother's maiden name. what type of microphone was actually used in that In an alternative embodiment of the invention, 10 case to produce the verification samples, a particular patterns are produced from unf iltered versions of the one of the four combinations would be expected to re- verification and reference samples and used along sult in the closest match, as also shown in the third with the patterns produced from the filtered versions column of the FIG. Given that all four possibilities are as described above. The unf iltered patterns are stor- available, there is no need to relax the matching cri- ed (412) in a portion of database 168 denoted by 15 teria, as in the prior art, in order to ensure that legiti- DB(412). As indicated above in steps 401,402, mate access attempts will be granted notwithstand- 404,406 and 407, the subscriber is prompted to recite ing the use of different microphone types. Presum- an utterance from which a filtered version of a refer- ably, one or another of the four possibilities will pro- ence pattern is ultimately produced and stored in da- duce a good match. tabase DB(407). 20 The theory under which the four-comparison ap- FIG. 6 represents the verification process which, proach may be more desirable than the single-com- for reasons that will become apparent, is referred to parison approach is that in the event that one of the as the four-way comparison process. An unfiltered patterns resulted from carbon-microphone-originat- version of the verification pattern is produced in steps ed samples, comparisons can be made which do not 601, 604 and 609 which parallels steps 401, 402 and 25 involve a double-filtered version of those samples. It 409, respectively, in FIG. 4. A filtered version of the also allows for a comparison between verification verification pattern is produced in steps 601, 604, patterns and reference patterns that were both pro- 606 and 611 which parallels steps 401, 402, 404 and duced from speech samples from electret micro- 406, respectively, in FIG. 4. (Although not explicitly phones, in which case filtering with a carbon-micro- shown, the prompting step 601 may include prompt- 30 phone characteristic should, at least in theory, not be ing the caller for the above-mentioned identification needed or helpful. Thus, the four-comparison ap- number.) In step 613, a four-way comparison is per- proach could, in theory, improve the overall system formed in which each version of the verification pat- performance. tern produced in steps 609 and 611 is compared to FIG. 7 is a conceptual view of the carbon-filter each version of the reference pattern retrieved from 35 process used in the flowcharts. The speech samples DB(407) and DB(412) (based on the subscriber iden- are operated on by a memoryless transfer function tification number). Ascore for each comparison is cal- 701. Transfer function 701 is non-linear, so that ener- culated as a function of the similarity of the charac- gy outside of the frequency band of interest-illustra- teristics of the verification patterns and reference tively 300-3300 Hz--may be created. This energy is patterns using well-known techniques. The compari- 40 removed by a bandpass filter 702. son having the "best" score, namely, the comparison FIG. 8 shows memoryless transfer function 701. having the highest degree of similarity, is chosen The FIG. shows, at any point in time, the value that is (617). Next, if it is determined that the two patterns output as a function of the instantaneous speech comprising the comparison having the best score are sample amplitude. The transfer function has three re- "close enough" to one another-based on well-known 45 gions: an expansion region, a linear region and a com- speech verification methods-access to the service pression region. The breakpoints Le and Lc and the provided by the platform is granted (624), service be- slopes Se and Sc are determined empirically. Specif- gins (633) and the process ends (635). ically, histograms of speech sample energy levels for If, on the other hand, the two patterns having the both carbon-microphone- and electret-microphone- best score are not "close enough" to one another, and 50 produced speech are obtained, and the four transfer this is not the caller's second attempt to gain access function parameters are selected based on those his- during this call, as determined at 622, the process re- tograms in such a way that the histogram of electret- turns to step 601 , thereby giving the caller a "second produced speech, once operated on by the transfer try." If it is the caller's second attempt, access is de- function, will match as nearly as possible the histo- nied (627) and the platform terminates the call. As dis- 55 gram of carbon-produced speech. cussed above, it may be possible that the caller is, in fact, the subscriber but for some reason access was denied. Accordingly, alternate methods may be used 5 g EP 0 654 781 A2 10

Claims 10. The method according to claim 9 wherein said comparing step further comprises the step of: 1. A method of processing speech samples for use comparing the filtered reference pattern in verifying a speaker, the method comprising the and the filtered verification pattern. steps of: 5 receiving speech samples generated by 11. The method according to claim 8 wherein said different types of microphones; generated reference pattern comprises a filtered filtering at least one of said speech sam- reference pattern produced by filtering the refer- ples generated by one of the microphone types ence speech sample with the characteristics of with the characteristics of another one of the mi- 10 said another microphone type, the method fur- crophone types; and ther comprising the steps of: producing patterns from said filtered producing an unfiltered verification pat- speech samples that can be used to verify the tern from a verification speech sample which has speaker. not been filtered; 15 producing an unfiltered reference pattern 2. The method according to claim 1 wherein said from a reference speech sample which has not speech samples are reference speech samples. been filtered; and comparing each verification pattern with 3. The method according to claim 1 wherein said each reference pattern. speech samples are verification speech sam- 20 ples. 12. The method according to claim 11 wherein said determining step further comprises the steps of: 4. The method according to claim 1 wherein the calculating a score representing the simi- characteristics of said another one of the micro- larity of the verification pattern and the reference phone types are non-linear. 25 pattern for each comparison; determining which comparison has the 5. The method according to claim 4 wherein said mi- best score based on predetermined criteria. crophone is a carbon microphone. 13. The method according to claim 8 wherein the 6. The method according to claim 1 wherein the 30 characteristics of said another microphone type characteristics of said one microphone type are are non-linear. linear. 14. The method according to claim 9 wherein the 7. The method according to claim 6 wherein said mi- characteristics of said another microphone type crophone is an electret microphone. 35 are non-linear.

8. A method of verifying a speaker, the method com- 15. The method according to claim 13 wherein said prising the steps of: microphone is a carbon microphone. receiving a speech sample provided for verification from one type of microphone; 40 16. The method according to claim 14 wherein said filtering the verification speech sample microphone is a carbon microphone. with the characteristics of another microphone type; 17. A method of processing speech samples for use producing a verification pattern from the in verifying a speaker, the method comprising the filtered verification speech sample; 45 steps of: comparing the verification pattern to a receiving speech samples impacted by generated reference pattern produced from a ref- factors characterized by different properties; erence speech sample provided during a regis- processing at least one of said speech tration session; and samples so that the properties characterizing the verifying the identity of the speaker as a so processed speech sample which are unrelated to function of the reference pattern and verification speaker differences are more similar to the prop- pattern. erties characterizing at least one of the other speech samples; and 9. The method according to claim 8 wherein said producing patterns from said processed generated reference pattern is produced by filter- 55 speech samples that can be used to verify the ing the reference speech sample with the charac- speaker. teristics of said another microphone type. 18. The method according to claim 17 wherein said 6 11 EP 0 654 781 A2 12

properties characterizing said one speech sam- ple are linear microphone characteristics.

19. The method according to claim 18 wherein said linear microphone characteristics are electret mi- 5 crophone characteristics.

20. The method according to claim 17 wherein said properties characterizing another one of said speech samples are non-linear microphone char- 10 acteristics.

21. The method according to claim 20 wherein said non-linear microphone characteristics are carbon microphone characteristics. 15

20

25

30

35

40

45

50

7 tP U b&4 /»1 Ait

JO r 12 14 15 LL 11 LI MICRO- OTHER RAM ROM PROCESSOR PERIPHERALS

10 J-

,11 19 111 TELEPHONE HAKU UISK 161 DSP INTERFACE MEMORY rA 1 1 bKn rKUUUUMUN 163 NLItKIINb KUUIINt 164 — ^21 PROGRAMS Ulntrc rKUbKAMS ILI LK UnAKAU tKloML "~ lb/ DATA 3V5V REFERENCEKLrLKLnLt loo PATTERNlAirrDki DATABASEnATADAcr UlnLK UAIAbAoLO

8 tr U D04 101 PU.

REFERENCE VERIFICATION CLOSEST MATCH PATTERN PATTERN WHEN MICROPHONES ARE:

»™ ™ FILTERED UNFILTERED

UNFILTERED FILTERED «j™ ^ s

UNFILTERED UNFILTERED gjj™ ELECTRET°R

B0TH CARB0N 0R FILTERED FILTEREDFll TFRFD gQTH ELECTRET

tlG. 4 KtlililKAIION

■'KUMrl

02

(ttllVt Ktrfc.RE.NCE SPEECH SAMPLES

,AKdUN HULK SAMPLES

Ub 09 PRODUCE !EFERENCE PATTERN IEFERENCE PATTERN

r ■ 07 12 — I ilUKt IN STORE IN ATABASE lATABASE

B (407) B (412) tr U D04 101 PU.

Mr. 0

VtKIMCATION

301

PROMPT

Ll KtUtlVL VERIFICATION SPEECH SAMPLE

)04 Ll L/AKdON FILTER SAMPLE

)07 Z_ rKUUULt VtKlrfCATION PATTERN

Ll ill 'B (407j LUMrAKt IU MLILKLO iTORED REFERENCE PATTERN

14 15 LL Ll rES ;LObt ENOUGH? >^ (_ 5EC0ND TRY

>17

GRANT ACCESS DENY

BEGIN SERVICE ) HANG UP ) . 24 \ _END_J tP U b&4 /»1 Ait

VERIFICATION (4-WAY)

,601

PROMPT

604

KtCEIVE VtRinCATION SPEECH SAMPLE

509 506 Z_ KKOUUCt tAKdUN VERIFICATION FILTER PATTERN SAMPLE

511 'KUUULt LAKoUN FILTERED VERIFICATION PATTERN /DB (407) _ . >13

4-WAY COMPARISON WITH TWO STORED REFERENCE PATTERNS

i17 Ll IB (412)

19 22

;L05E ENOUGH? >s <_ SECOND TRY

|YES E5 ,624 r w 27 — i GRANT ACCESS )ENY

/633 r n 31 / — Ir-^ v BEGIN SERVICE ) END CALL) ^ , 35 f " _END_J

i EP 0 654 781 A2

FIG. 7

,701 r 702 INFILE MEMORYLESS BAND-PASS OUTFILE NONLINEARITY FILTER

FIG. 8

OUT

REGION 3 REGION 2 REGION 1 REGION 2 REGION 3

12