Method and apparatus for training a text independent speaker recognition system using speech data with text labels

• US 7,813,927 B2
• Filed: 06/04/2008
• Issued: 10/12/2010
• Est. Priority Date: 11/22/2004
• Status: Active Grant

• Current Assignee: Microsoft Technology Licensing LLC (Microsoft Corporation)
• Original Assignee: Nuance Communications, Inc. (Microsoft Corporation)
• Inventors: Nealand, James H., Navratil, Jiri, Pelecanos, Jason W., Ramaswamy, Ganesh N., Zilca, Ran D.
• Primary Examiner(s): Dorvil; Richemond
• Assistant Examiner(s): GODBOLD, DOUGLAS
• Application Number:
  US12/132,770
• Publication Number:
  US 20080235020A1
• Time in Patent Office: 860 Days
• Field of Search: 704246-250, 704/273
• US Class Current: 704/250
• CPC Class Codes: G10L 15/063 Training
• CPC Class Codes: G10L 15/144 Training of HMMs

Abstract

There is provided an apparatus for providing a Text Independent (TI) speaker recognition mode in a Text Dependent (TD) Hidden Markov Model (HMM) speaker recognition system and/or a Text Constrained (TC) HMM speaker recognition system. The apparatus includes a Gaussian Mixture Model (GMM) generator and a Gaussian weight normalizer. The GMM generator is for creating a GMM by pooling Gaussians from a plurality of HMM states. The Gaussian weight normalizer is for normalizing Gaussian weights with respect to the plurality of HMM states.

Claims

1. A Hidden Markov Model (HMM) speaker recognition system, comprising:

at least one data storage unit; and

at least one processor programmed to;

create a Gaussian Mixture Model (GMM) by pooling Gaussians from a plurality of HMM states; and

normalize Gaussian weights with respect to the plurality of HMM states to provide a Text Independent (TI) speaker recognition mode in the HMM speaker recognition system;

wherein the HMM speaker recognition system is selected from a group consisting of a Text Dependent (TD) HMM speaker recognition system and a Text Constrained (TC) speaker recognition system.

2. The system of claim 1, wherein said at least one processor is further programmed to:

in creating the GMM, derive a GMM Universal Background Model (UBM) from an HMM UBM and/or derive a speaker specific GMM from a speaker specific HMM.

3. The system of claim 1, wherein said at least one processor is further programmed to:

normalize the Gaussian weights based on durations of the plurality of HMM states and/or normalize the Gaussian weights by dividing each of the Gaussian weights by an overall number of the plurality of HMM states.

4. The system of claim 1, wherein the HMM speaker recognition system is based on a single HMM per word, a single HMM per phonetic unit, and/or a single HMM per sub-phonetic unit.

5. A Text Independent (TI) Gaussian Mixture Model (GMM) speaker recognition system, comprising:

at least one data storage unit; and

at least one processor programmed to;

create a Hidden Markov Model (HMM) by assigning states to Gaussians from a GMM; and

calculate state transition probabilities and Gaussian weights with respect to a plurality of HMM states to provide a Text Dependent (TD) HMM speaker recognition mode and/or a Text Constrained (TC) HMM speaker recognition mode in the TI GMM speaker recognition system.

6. The system of claim 5, wherein said at least one processor is further programmed to:

in creating the HMM, derive a HMM Universal Background Model (UBM) from a GMM UBM and/or derive a speaker specific HMM from a speaker specific GMM.

7. The system of claim 5, wherein said at least one processor is further programmed to:

calculate the state transition probabilities and the Gaussian weights based on durations of the plurality of HMM states and/or based on only an overall number of the plurality of HMM states.

8. The system of claim 5, wherein the Text Dependent (TD) Hidden Markov Model (HMM) speaker recognition mode and/or the Text Constrained (TC) HMM speaker recognition mode is based on a single HMM per word, a single HMM per phonetic unit, and/or a single HMM per sub-phonetic unit.

9. A first Hidden Markov Model (HMM) speaker recognition system, comprising:

at least one data storage unit; and

at least one processor programmed to;

create an HMM with a smaller number of states or a larger number of states by pooling Gaussians from a plurality of HMM states into a single HMM state or splitting the Gaussians from the plurality of HMM states into different HMM states, respectively; and

normalize Gaussian weights with respect to the HMM states to provide a TD HMM speaker recognition mode and/or a TC HMM speaker recognition mode in a second HMM speaker recognition system;

wherein the first HMM speaker recognition system is selected from a group consisting of a Text Dependent (TD) HMM speaker recognition system and a Text Constrained (TC) speaker recognition system.

10. The system of claim 9, wherein said at least one processor is further programmed to:

in creating the HMM, derive a first HMM Universal Background Model (UBM) from a second HMM UBM and/or derive a first speaker specific HMM from a second speaker specific HMM.

11. The system of claim 9, wherein said at least one processor is further programmed to:

normalize the Gaussian weights based on HMM state durations and/or normalize the Gaussian weights by dividing each of the Gaussian weights by an overall number of the HMM states.

12. The system of claim 9, wherein the second HMM speaker recognition system includes a single HMM per phoneme and/or a single HMM per phonetic unit, and the TD HMM speaker recognition mode and/or the TC HMM speaker recognition mode includes a single HMM per word.

13. The system of claim 9, wherein the second HMM speaker recognition system includes a single HMM per word and the TD HMM speaker recognition mode and/or the TC HMM speaker recognition mode includes a single HMM per phrase.