Policy analysis framework for conversational biometrics

• US 7,363,223 B2
• Filed: 08/13/2004
• Issued: 04/22/2008
• Est. Priority Date: 08/13/2004
• Status: Expired due to Fees

• Current Assignee: Google LLC (Alphabet Inc.)
• Original Assignee: International Business Machines Corporation
• Inventors: Ramaswamy, Ganesh N., Chaudhari, Upendra V.
• Primary Examiner(s): Edouard; Patrick N.
• Assistant Examiner(s): Wozniak; James S.
• Application Number:
  US10/917,992
• Publication Number:
  US 20060036443A1
• Time in Patent Office: 1,348 Days
• Field of Search: 704246-250, 726 5- 6
• US Class Current: 704/246
• CPC Class Codes: G10L 17/22 Interactive procedures; Man-machine interfaces

Abstract

The present invention describes a framework for the analytical and visual analysis and tuning of complex speaker verification systems guided by a policy finite state machine. The Receiver Operating Curve associated with the acoustic speaker recognition task is transformed into a multi-dimensional Receiver Operating Map (ROM), which results from a probabilistic analysis of the policy state machine. A Detection Cost Function (DCF) Map can be similarly generated. Results indicating that optimization over this surface (or the ROM) is an appropriate way to set thresholds are given.

Claims

1. A method of assessing error rates and tuning parameters associated with policy based procedures for verification of a plurality of known speakers through use of conversational biometrics, evolution of the policy based procedures being controlled by a state machine and each of the plurality of known speakers has an associated acoustic component and a knowledge profile, the method comprising the steps of:

performing a policy based verification using a finite state machine defined by a set of states, each state having a specified set of questions and a plurality of possible state transitions, a state transition being performed if a condition of a state transition path associated with the state transition is satisfied;

determining allowability of the condition;

generating a state transition path through the state machine;

identifying all allowable state transition paths;

developing a probabilistic analysis of the behavior of the state machine, wherein each allowable state transition path is assigned a probability value;

generating a Receiver Operating Map (ROM) by using said probability values;

mapping a set of thresholds/parameters to a false accept rate and a false reject rate by using a multidimensional acoustic Receiver Operating Curve and probabilities of knowledge error, wherein the ROM is a function of a plurality of the thresholds/parameters with an output selected from one of accept rate and reject rate; and

tuning said thresholds/parameters by visually or analytically examining and optimizing over the ROM.

2. The method of claim 1, wherein allowability of the condition is determined by transforming a plurality of conditions into a set of linear constraints or variables and determining a feasible set.

3. The method of claim 1, wherein the performing step further includes analyzing speech input of said known speakers with respect to the acoustic component and the knowledge profile of the speech input.

4. The method of claim 3, wherein the acoustic component is based on a likelihood ratio statistic and the knowledge profile is represented as a table of questions and answers that are specific to each of the plurality of said known speakers.

5. The method of claim 1, wherein the condition includes a plurality of variables and a plurality of constants, the plurality of constants including three acoustic score thresholds/parameters:

low (l), medium (m), and high (h) and the plurality of variables include

T, a number of topics covered before transition path conditions are assessed,

W, a number of topics covered for which the answer given was incorrect, and

S, a current acoustic score threshold/parameter.

6. The method of claim 1, wherein the step of generating a state transition path further comprises the steps of:

a) asking a question associated with a state, the question have not been asked before on the state transition path,

b) obtaining a response to the asked question and a response score threshold/parameter,

c) processing the conditions until one condition is satisfied;

d) performing the state transition corresponding to the condition; and

e) repeating steps a) to d) if a present state is not a terminal state and otherwise, making a decision of generating an ACCEPT or REJECT state.

7. The method of claim 6, wherein the step of generating a state transition path is simulated.

8. The method of claim 1, wherein for the step of identifying is determined according to whether a full linear or mixed integer program of the state transition path is solvable.

9. The method of claim 1, wherein the probability value of the accept rate is a probability of transitioning to an ACCEPT state and the probability value of the reject rate is a probability of transitioning to a REJECT state, the probability value of the reject rate is indicated as 1−

accept rate.

10. The method of claim 9, wherein the probability of transitioning to an ACCEPT state is the probability of all state transition paths leading to the ACCEPT state and the probability of transitioning to a REJECT state is the probability of all paths leading to the REJECT state.

11. The method of claim 10, wherein the step of generating a ROM further comprises the steps of:

defining a grid of threshold/parameter values in a plurality of dimensions, wherein one dimension is allocated to each of low, medium, and high score threshold/parameters;

computing the probability value of the accept rate, or probability value of the reject rate for each set of threshold/parameter values given by the grid;

creating a visual representation of the accept or reject rates as a function of the threshold/parameter values given by the grid; and

creating a representation of the accept or reject rates as a function of the threshold/parameter values given by the grid.

12. The method of claim 11, wherein an upper surface of the ROM corresponds to a correct speaker of the plurality of known speakers, a middle surface corresponds to informed impersonators who have gained access to knowledge associated with the known speaker, and the lowest surface corresponds to uninformed impersonators who do not have knowledge associated with the known speaker.

13. The method of claim 11, wherein the tuning step further comprises the step of determining one or more desirable points in the ROM, based on the probabilities of accept rate and reject rate for target known speakers and one or more impersonators of known speakers, wherein said determining is performed using visual inspection of the ROM.

14. The method of claim 11, wherein the tuning step further comprises the steps of:

evaluating regions of the ROM using analytical methods; and

determining one or more desirable points in the ROM based on the probabilities of accept rate and reject rate for target known speakers one or more impersonators of known speakers, wherein said determining is performed via analytical optimization.

15. The method of claim 14, wherein the analytical optimization maximizes an objective function.

16. The method of claim 14, wherein the analytical optimization minimizes an objective function.

17. A method of assessing error rates and tuning parameters associated with policy based procedures for verification of a plurality of known speakers through use of conversational biometrics, evolution of the policy based procedures being controlled by a state machine and each of the plurality of known speakers has an associated acoustic component and a knowledge profile, the method comprising the steps of:

performing a policy based verification using a finite state machine defined by a set of states, each state having a specified set of questions and a plurality of possible state transitions, a state transition being performed if a condition of a state transition path associated with the state transition is satisfied;

generating a Receiver Operating Map (ROM) by using a plurality of probability values;

mapping a set of thresholds/parameters to a false accept rate and a false reject rate by using a multidimensional acoustic Receiver Operating Curve and probabilities of knowledge error, wherein the ROM is a function of a plurality of the thresholds/parameters with an output selected from one of accept rate and reject rate; and

utilizing the mapped set of thresholds/parameters in a speaker recognition process.

18. The method of claim 17, further comprising the steps of:

developing a probabilistic analysis of the behavior of the state machine, wherein each allowable state transition path is assigned a probability value;

determining allowability of the condition;

generating a state transition path through the state machine; and

identifying all allowable state transition paths.

19. The method of claim 17, further comprising a step of tuning said thresholds/parameters by examining and optimizing said thresholds/parameters over the ROM, wherein said examining and optimizing is performed by a method selected from a visual and analytical.

20. The method of claim 17, wherein the condition includes a plurality of variables and a plurality of constants.