Methods and system for reducing false positive voice print matching

• US 10,410,636 B2
• Filed: 12/01/2017
• Issued: 09/10/2019
• Est. Priority Date: 11/09/2012
• Status: Active Grant

• Current Assignee: Mattersight Corporation (Nice Ltd)
• Original Assignee: Mattersight Corporation (Nice Ltd)
• Inventors: Warford, Roger, Brown, Douglas, Danson, Christopher, Gustafson, David
• Primary Examiner(s): Sharma, Neeraj
• Application Number:
  US15/829,529
• Publication Number:
  US 20180082690A1
• Time in Patent Office: 648 Days
• Field of Search: None
• 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 2025/783 based on threshold decision
• CPC Class Codes: G10L 25/27 characterised by the analysis technique
• CPC Class Codes: G10L 25/51 for comparison or discrimination
• CPC Class Codes: G10L 25/78 Detection of presence or absence of voice signals switching of direction of transmission by voice frequency in two-way loud-speaking telephone systems H04M9/10

Abstract

The methods, apparatus, and systems described herein are designed to reduce false positive voice print matching with fraudulent callers. A voice print of a call is created and compared to known voice prints to determine if it matches one or more of the known voice prints, and to transaction data associated with a database of voice prints. The methods include a pre-processing step to separate speech from non-speech, selecting a number of elements that affect the voice print the most, and/or generating a first score based on the number of selected audio elements matching audio elements of a voice print from the plurality of fraudulent speakers, determining if the first score exceeds a predetermined threshold score for the fraudulent speaker, and comparing the selected audio elements for the unknown caller, where the score exceeds the predetermined threshold score, to the voice prints associated with the customer account.

Claims

1. A method of reducing false positive matches in voice prints which comprises:

receiving an audio communication from an unknown caller, separating a first portion of the audio communication into silent and non-silent segments, and evaluating the non-silent segments to determine which portions thereof are speech or non-speech;

generating a plurality of parameters that determine what is speech and non-speech in the non-silent segments;

using the generated parameters to determine what is speech and non-speech for at least the remainder of the audio communication;

comparing the speech to selected audio elements of a universal background model (UBM) that characterizes the speech of the unknown caller relative to a plurality of other audio elements of the UBM;

comparing the selected audio elements of the speech to matching audio elements of a recorded voice print from a plurality of fraudulent speakers to determine whether the speech belongs to a fraudulent speaker;

generating a first score based on the number of selected audio elements matching audio elements of the fraudulent speaker from the plurality of fraudulent speakers;

determining if the first score exceeds a predetermined threshold score for the fraudulent speaker; and

comparing the selected audio elements for the unknown caller where the score exceeds the predetermined threshold score to the selected audio elements of a voice print associated with a customer account and generating a second score based on the number of selected audio elements matching audio elements associated with the customer account;

wherein if the value of the first score exceeds the value of the second score, then generating an alert for display on an agent workstation.

2. The method of claim 1, which further comprises routing the unknown caller to an agent or a manager assigned to handle fraudulent callers.

3. The method of claim 1, wherein the predetermined threshold score is based on a dynamic threshold and is reduced if the caller has a first score that is greater than zero, thereby indicating a possible fraudulent caller.

4. The method of claim 1, wherein the plurality of parameters are generated for each communication received.

5. The method of claim 1, wherein evaluating the non-silent segments further comprise instructions that, when executed, treat all non-speech sounds substantially the same.

6. A non-transitory computer readable medium comprising a plurality of instructions stored therein that are adapted to reduce false positive matches when analyzing voice prints, the plurality of instructions comprising:

instructions that, when executed, receive an audio communication from an unknown caller, separating a first portion of the audio communication into silent and non-silent segments, and evaluating the non-silent segments to determine which portions thereof are speech or non-speech;

instructions that, when executed, generate a plurality of parameters that determine what is speech and non-speech in the non-silent segments;

instructions that, when executed, use the generated parameters to determine what is speech and non-speech for at least the remainder of the audio communication;

instructions that, when executed, compare the speech to selected audio elements of a universal background model (UBM) that characterizes the speech of the unknown caller relative to a plurality of other audio elements of the UBM;

instructions that, when executed, compare the selected audio elements of the speech to matching audio elements of a recorded voice print from a plurality of fraudulent speakers to determine whether the speech belongs to a fraudulent speaker;

instructions that, when executed, generate a first score based on the number of selected audio elements matching audio elements of the fraudulent speaker from the plurality of fraudulent speakers;

instructions that, when executed, determine if the first score exceeds a predetermined threshold score for the fraudulent speaker; and

instructions that, when executed, compare the selected audio elements for the unknown caller where the score exceeds the predetermined threshold score to the selected audio elements of a voice print associated with a customer account and generating a second score based on the number of selected audio elements matching audio elements associated with the customer account;

wherein if the value of the first score exceeds the value of the second score, then instructions that, when executed, generate an alert for display on an agent workstation.

7. The computer-readable medium of claim 6, which further comprises instructions that, when executed, route the unknown caller to an agent or a manager assigned to handle fraudulent callers.

8. The computer-readable medium of claim 6, wherein the predetermined threshold score is based on a dynamic threshold and is reduced if the caller has a first score that is greater than zero, thereby indicating a possible fraudulent caller.

9. The computer-readable medium of claim 6, wherein the plurality of parameters are generated for each communication received.

10. The computer-readable medium of claim 6, wherein the instructions that, when executed, evaluate the non-silent segments further comprise instructions that, when executed, treat all non-speech sounds substantially the same.

11. An audible fraud detection system adapted to minimize false positive voice print matches, which comprises:

a node comprising a processor and a non-transitory computer readable medium operably coupled thereto, the non-transitory computer readable medium comprising a plurality of instructions stored therein and being accessible to, and executable by, the processor, where the plurality of instructions comprises;

instructions that, when executed, receive an audio communication from an unknown caller, separating a first portion of the audio communication into silent and non-silent segments, and evaluating the non-silent segments to determine which portions thereof are speech or non-speech;

instructions that, when executed, generate a plurality of parameters that determine what is speech and non-speech in the non-silent segments;

instructions that, when executed, use the generated parameters to determine what is speech and non-speech for at least the remainder of the audio communication;

instructions that, when executed, compare the speech to selected audio elements of a universal background model (UBM) that characterizes the speech of the unknown caller relative to a plurality of other audio elements of the UBM;

instructions that, when executed, compare the selected audio elements of the speech to matching audio elements of a recorded voice print from a plurality of fraudulent speakers to determine whether the speech belongs to a fraudulent speaker;

instructions that, when executed, generate a first score based on the number of selected audio elements matching audio elements of the fraudulent speaker from the plurality of fraudulent speakers;

instructions that, when executed, determine if the first score exceeds a predetermined threshold score for the fraudulent speaker; and

instructions that, when executed, compare the selected audio elements for the unknown caller where the score exceeds the predetermined threshold score to the selected audio elements of a voice print associated with a customer account and generating a second score based on the number of selected audio elements matching audio elements associated with the customer account;

wherein if the value of the first score exceeds the value of the second score, then instructions that, when executed, generate an alert for display on an agent workstation.

12. The system of claim 11, wherein the number of selected elements is from about 10 to 30.

13. The system of claim 11, further comprising instructions that, when executed, identify the gender of the voice print.

14. The system of claim 11, wherein the recorded voice prints from a plurality of fraudulent speakers are stored in a database.

15. The system of claim 11, which further comprises routing the unknown caller to an agent or a manager assigned to handle fraudulent callers.

16. The system of claim 11, wherein the predetermined threshold score is based on a dynamic threshold and is reduced if the caller has a first score that is greater than zero, thereby indicating a possible fraudulent caller.

17. The system of claim 11, wherein the plurality of parameters are generated for each communication received.

18. The system of claim 11, wherein the instructions that, when executed, evaluate the non-silent segments further comprise instructions that, when executed, treat all non-speech sounds substantially the same.