Joint factor analysis scoring for speech processing systems

1. A computer-implemented method for Joint Factor Analysis (JFA) scoring in speech processing systems performed by a computerized device, comprising:

carrying out an enrollment session offline to enroll a speaker model in a speech processing system using JFA, including:

providing an equation for calculating a score representative of the extent to which a speaker of a received test speech signal matches an enrolled speaker, based on speaker factors, a first component of channel factors which loosely depend on the test speech signal and a second component of channel factors which strongly depend on the test speech signal;

receiving an enrollment speech signal from a user;

extracting the speaker factors from the enrollment speech signal;

estimating one or more values of the first component of channel factors from the enrollment speech signal, without using a test speech signal;

carrying out a test session including:

receiving a test speech signal;

retrieving an estimated first component of channel factors which was estimated in the enrollment session without using the test speech signal, for the received test speech signal;

calculating the second component of channel factors strongly dependent on the test session, based on the test speech signal; and

generating a score based on the provided equation, the speaker factors, the estimated first component and the calculated second component of channel factors.

2. The method as claimed in claim 1 , wherein estimating one or more values of the first component of channel factors is carried out separately for a plurality of genders, based on a plurality of different enrollment speech signals belonging to speakers of different genders.

3. The method as claimed in claim 1 , wherein estimating one or more values of the first component of channel factors is carried out separately for a plurality of different durations of sessions, based on a plurality of different enrollment speech signals of different durations.

4. The method as claimed in claim 1 , wherein the speech processing is speaker verification and generating a score calculates a log-likelihood ratio between a target speaker model and the universal background model for a test session.

5. The method as claimed in claim 1 , wherein the first component of the channel factors generalizes the role of the relevance factor used in a universal background model, Gaussian mixture model maximum a posteriori (UBM-GMM MAP) adaptation.

6. The method as claimed in claim 1 , wherein the first and second components are components needed to calculate the posterior expectation of factors.

7. The method as claimed in claim 1 , wherein:

estimating one or more values of the first component of channel factors comprises calculating the zero order statistics (matrix N) for a plurality of enrollment speech signals;

calculating the mean of the calculated zero order statistics (matrix N); and

using the mean of matrix N for calculating the first components.

8. The method as claimed in claim 1 , estimating one or more values of the first component of channel factors comprises

estimating first components, β and ζ −1 of a linear scoring method equation by computing the first components β T and T 1 for a plurality of enrollment speech signals and calculating their means.

9. The method as claimed in claim 8 , wherein the test session includes:

calculating the zero and first order statistics (F, N) for the enrollment speech;

calculating the second component of the channel factors γ −1 using N; and

estimating channel factors using the second component γ −1 , the zero and first order statistics F, N, and the pre-calculated first components β and ζ −1 .

10. A computer program product for Joint Factor Analysis (JFA) scoring in speech processing systems, embodied in a computer readable non-transitory storage medium the computer readable program product comprising:

computer readable program code configured to:

carry out an enrollment session offline to enroll a speaker model in a speech processing system using JFA, including:

providing an equation for calculating a score representative of the extent to which a speaker of a received test speech signal matches an enrolled speaker, based on speaker factors, a first component of channel factors which loosely depend on the test speech signal and a second component of channel factors which strongly depend on the test speech signal;

receiving an enrollment speech signal from a user;

extracting the speaker factors from the enrollment speech signal;

estimating one or more values of the first components component of channel factors from the enrolment enrollment speech signal, without using a test speech signal session; and

carry out a test session including:

receiving a test speech signal;

retrieving an estimated first component of channel factors which was estimated in the enrollment session without using the test speech signal, for the received test speech signal;

calculating the second component of channel factors strongly dependent on the test session, based on the test speech signal; and

generating a score based on the provided equation, the speaker factors, the estimated first component and the calculated second component of channel factors.

11. A system for Joint Factor Analysis (JFA) scoring in a speech processing system, comprising:

a processor;

an enrollment system for carrying out an enrollment session offline to enroll a speaker model in a speech processing system using JFA, including:

providing an equation for calculating a score representative of the extent to which a speaker of a received test speech signal matches an enrolled speaker, based on speaker factors, a first component of channel factors which loosely depend on the test speech signal and a second component of channel factors which strongly depend on the test speech signal;

receiving an enrollment speech signal from a user;

extracting the speaker factors from the enrollment speech signal; and

estimating one or more values of the first component of channel factors from the enrollment speech signal, without using a test speech signal; and

a test system for carrying out a test session including:

receiving a test speech signal;

retrieving an estimated first component of channel factors which was estimated in the enrollment session without using the test speech signal, for the received test speech signal;

calculating the second component of channel factors strongly dependent on the test session, based on the test speech signal; and

generating a score based on the provided equation, the speaker factors, the estimated first component and the calculated second component of channel factors.

12. The method as claimed in claim 1 , wherein estimating one or more values of the first component of channel factors comprises computing the first components for a plurality of enrollment speech signals and calculating a mean of the computed first components.

13. The method of claim 1 , wherein the equation is a test session Gaussian mixture model equation sufficient to provide a log-likelihood ratio for a test session.

14. The method of claim 1 , wherein providing the equation comprises providing an equation in accordance with a linear scoring method.