System and method for generating challenge items for CAPTCHAs

• US 8,380,503 B2
• Filed: 06/15/2009
• Issued: 02/19/2013
• Est. Priority Date: 06/23/2008
• Status: Active Grant

• Current Assignee: Knapp Investment Company Limited
• Original Assignee: John Nicholas and Kristin Gross Trust
• Inventors: Gross, John Nicholas
• Primary Examiner(s): Dorvil, Richemond
• Assistant Examiner(s): ADESANYA, OLUJIMI A
• Application Number:
  US12/484,870
• Publication Number:
  US 20090319271A1
• Time in Patent Office: 1,345 Days
• Field of Search: 704/246
• US Class Current: 704/246
• CPC Class Codes: G06F 21/32 using biometric data, e.g. fingerprints, iris scans or voiceprints
• CPC Class Codes: G10L 13/027 Concept to speech synthesisers; Generation of natural phrases from machine-based concepts generation of parameters for speech synthesis out of text G10L13/08
• CPC Class Codes: G10L 13/08 Text analysis or generation of parameters for speech synthesis out of text, e.g. grapheme to phoneme translation, prosody generation or stress or intonation determination
• CPC Class Codes: G10L 15/02 Feature extraction for speech recognition; Selection of recognition unit
• CPC Class Codes: G10L 15/063 Training
• CPC Class Codes: G10L 15/22 Procedures used during a speech recognition process, e.g. man-machine dialogue
• CPC Class Codes: G10L 17/00 Speaker identification or verification techniques
• CPC Class Codes: G10L 17/02 Preprocessing operations, e.g. segment selection; Pattern representation or modelling, e.g. based on linear discriminant analysis [LDA] or principal components; Feature selection or extraction
• CPC Class Codes: G10L 17/04 Training, enrolment or model building
• CPC Class Codes: G10L 17/06 Decision making techniques; Pattern matching strategies
• CPC Class Codes: G10L 17/22 Interactive procedures; Man-machine interfaces
• CPC Class Codes: G10L 17/26 Recognition of special voice characteristics, e.g. for use in lie detectors; Recognition of animal voices
• CPC Class Codes: H04L 63/102 Entity profiles
• CPC Class Codes: H04L 63/123 received data contents, e.g. message integrity
• CPC Class Codes: H04M 2203/2027 Live party detection

Abstract

Challenge items for an audible based electronic challenge system are generated using a variety of techniques to identify optimal candidates. The challenge items are intended for use in a computing system that discriminates between humans and text to speech (TTS) system.

Claims

1. A method embodied in a computer readable medium for generating challenge data to be used for accessing data and/or resources of an electronic computing system comprising:

(a) automatically generating a candidate challenge sentence from a first set of words and phrases using a computing system;

(b) automatically generating at least one first utterance from a first machine text to speech system of said candidate challenge sentence using the computing system, said at least one first utterance including first acoustical characteristics;

(c) acquiring at least one second utterance of said candidate challenge sentence known to be from a human speaker using the computing system, said at least one second utterance including second acoustical characteristics;

(d) automatically determining a difference in said first and second acoustical characteristics using the computing system to determine a challenge sentence acoustic score for said candidate challenge sentence;

(e) automatically selecting said candidate challenge sentence as a suitable challenge item in an utterance-based challenge system and storing said at least one second utterance and said candidate challenge sentence in a challenge item database using the computer system when said candidate challenge sentence acoustic score exceeds a target threshold, indicating that said at least one first utterance from the first machine text to speech system is sufficiently different from said at least one second utterance from the human speaker for said candidate challenge sentence to be used to automatically distinguish between humans and machines attempting to access the electronic computing system.

2. The method of claim 1, wherein acoustical features of said first utterance and said second utterance are measured, and regions of greatest difference are identified to be used as discriminators.

3. The method of claim 1 wherein differences in articulation are measured and scored on a diphone basis.

4. The method of claim 3 wherein said diphones are sorted according to their difference between human and machine articulation.

5. The method of claim 3 further including a step:

selecting words and phrases from a corpus of text in accordance with a diphone score.

6. The method of claim 1 further including a step:

compiling a machine articulation statistical language model based on a plurality of first utterances.

7. The method of claim 1 further including a step:

compiling a dialog of multiple challenge sentences based on questions and expected answers to be provided by an entity.

8. The method of claim 1 further including a step:

generating a challenge scenario comprised of said candidate challenge sentence and one or more visual and/or text cues, which challenge scenario is stored in said database.

9. The method of claim 1 further including a step:

processing said words and phrases with a natural language engine to generate said candidate challenge sentence.

10. The method of claim 1 further including a step:

annotating said candidate challenge sentence with prosodic elements to generate an expected prosodic pronunciation of said words and phrases.

11. The method of claim 8 further including a step:

annotating said candidate challenge sentence with first prosodic elements found in said at least one first utterance and with second prosodic elements found in said least one second utterance.

12. The method of claim 11 further including a step:

determining a difference in said first and second prosodic elements to determine a challenge sentence prosodic score for said candidate challenge sentence.

13. The method of claim 8 further including a step:

annotating said candidate challenge sentence with first content elements for said visual and/or text cues found in said at least one first utterance and with second content elements for said visual and/or text cues found in said least one second utterance.

14. The method of claim 12 further including a step:

determining a difference in said first and second content elements to determine a challenge sentence content score for said candidate challenge sentence.

15. The method of claim 1 wherein a time required by the human speaker to generate said first utterance is measured.

16. The method of claim 1 wherein a time required by the machine speaker to generate said second utterance is measured.

17. The method of claim 1 where said first set of words and phrases are automatically extracted from an electronic corpus.

18. The method of claim 1 wherein said first set of words and phrases include homographs.

19. The method of claim 1 further including a step:

processing input speech by an entity using said challenge item database to distinguish between a human and a machine synthesized voice.

20. The method of claim 1, wherein said challenge sentence is associated with one or both of one or more visual images or one or more topics and said challenge sentence score is also based on one or both of the one or more visual images or the one or more topics identified and measured in said first and second utterances, respectively.

21. The method of claim 20, wherein said challenge sentence score is also based on a collaborative filtering score generated by measuring responses to a sequence of two or more of said candidate challenge sentences identified in said first and second utterances, respectively.

22. The method of claim 21 wherein said collaborative filtering score is derived by identifying at least a first reference correlation between at least a first reference text descriptor for a first challenge sentence and a second reference text descriptor for a second challenge sentence, including a probability that a human reviewer identifying said first reference text descriptor when presented with said first challenge sentence also provides said second reference text descriptor when presented with said second challenge sentence, or that the human reviewer identifying said second reference text descriptor when presented with said second challenge sentence provides said first reference text descriptor when presented with said first challenge sentence.

23. The method of claim 21 wherein said collaborative filtering score is derived by identifying at least a first reference correlation between a first challenge sentence presented in incomplete form, and a predicted response for completing said challenge sentence.

24. A system for identifying challenge data to be used for distinguishing between humans and machines in allowing access to computing resources, comprising:

a computing system; and

one or more software routines embodied in a computer readable medium adapted to cause the computing system to;

(a) automatically generate a candidate challenge sentence from a first set of words and phrases;

(b) automatically generate at least one first utterance from a first machine text to speech system of said candidate challenge sentence, said at least one first utterance including first acoustical characteristics;

(c) acquire at least one second utterance of said candidate challenge sentence known to be from a human speaker, said at least one second utterance including second acoustical characteristics;

(d) automatically determine a difference in said first and second acoustical characteristics to determine a challenge sentence acoustic score for said candidate challenge sentence;

(e) automatically select said candidate challenge sentence as a suitable challenge item in an utterance-based challenge system and store said at least one second utterance and said candidate challenge sentence in a challenge item database using the computer system when said candidate challenge sentence acoustic score exceeds a target threshold, indicating that said at least one first utterance from the first machine text to speech system is sufficiently different from said at least one second utterance from the human speaker for said candidate challenge sentence to be used to automatically distinguish between humans and machines.

25. The system of claim 24, wherein acoustical features of said first utterance and said second utterance are measured, and regions of greatest difference are identified to be used as discriminators.

26. The system of claim 24 wherein differences in articulation are measured and scored on a diphone basis.

27. The system of claim 26 wherein said diphones are sorted according to their difference between human and machine articulation.

28. The system of claim 26 wherein the one or more software routines also cause the computing system to select words and phrases and selected from a corpus of text in accordance with a diphone score.

29. The system of claim 24 wherein the one or more software routines also cause the computing system to compile a machine articulation statistical language model based on a plurality of first utterances.

30. The system of claim 24 wherein the one or more software routines also cause the computing system to compile a dialog of multiple challenge sentences based on questions and expected answers to be provided by an entity.

31. The system of claim 24 wherein the one or more software routines also cause the computing system to generate a challenge scenario comprised of said candidate challenge sentence and one or more visual and/or text cues, which challenge scenario is stored in said database.

32. The system of claim 31 wherein the one or more software routines also cause the computing system to annotate said candidate challenge sentence with first prosodic elements found in said at least one first utterance and with second prosodic elements found in said least one second utterance.

33. The system of claim 32 wherein the one or more software routines also cause the computing system to determine a difference in said first and second prosodic elements to determine a challenge sentence prosodic score for said candidate challenge sentence.

34. The system of claim 31 wherein the one or more software routines also cause the computing system to annotate said candidate challenge sentence with first content elements for said visual and/or text cues found in said at least one first utterance and with second content elements for said visual and/or text cues found in said least one second utterance.

35. The system of claim 34 wherein the one or more software routines also cause the computing system to determine a difference in said first and second content elements to determine a challenge sentence content score for said candidate challenge sentence.

36. The system of claim 24 wherein the one or more software routines also cause the computing system to process said words and phrases with a natural language engine to generate said candidate challenge sentence.

37. The system of claim 24 wherein the one or more software routines also cause the computing system to annotate said candidate challenge sentence with prosodic elements to generate an expected prosodic pronunciation of said words and phrases.

38. The system of claim 24 wherein a time required by the human speaker to generate said first utterance is measured.

39. The system of claim 24 wherein a time required by the machine speaker to generate said second utterance is measured.

40. The system of claim 24 where said first set of words and phrases are automatically extracted from an electronic corpus.

41. The system of claim 24 wherein said first set of words and phrases include homographs.

42. The system of claim 24 wherein the one or more software routines also cause the computing system to process input speech by an entity using said challenge item database to distinguish between a human and a machine synthesized voice.