SYSTEM AND METHOD FOR GENERATING USER MODELS FROM TRANSCRIBED DIALOGS

• US 20110054893A1
• Filed: 09/02/2009
• Published: 03/03/2011
• Est. Priority Date: 09/02/2009
• Status: Active Grant

• Current Assignee: Microsoft Technology Licensing LLC (Microsoft Corporation)
• Original Assignee: AT&T Intellectual Property I LP (AT&T, Inc.)
• Inventors: Williams, Jason, Syed, Umar
• Granted Patent: US 8,473,292 B2
• Time in Patent Office: Days
• US Class Current: 704/235
• CPC Class Codes: G10L 15/07 to the speaker
• CPC Class Codes: G10L 15/26 Speech to text systems G10L15/08 takes precedence
• CPC Class Codes: G10L 2015/0631 Creating reference templates; Clustering

Abstract

Disclosed herein are systems, computer-implemented methods, and computer-readable storage media for generating personalized user models. The method includes receiving automatic speech recognition (ASR) output of speech interactions with a user, receiving an ASR transcription error model characterizing how ASR transcription errors are made, generating guesses of a true transcription and a user model via an expectation maximization (EM) algorithm based on the error model and the respective ASR output where the guesses will converge to a personalized user model which maximizes the likelihood of the ASR output. The ASR output can be unlabeled. The method can include casting speech interactions as a dynamic Bayesian network with four variables: (s), (u), (r), (m), and encoding relationships between (s), (u), (r), (m) as conditional probability tables. At each dialog turn (r) and (m) are known and (s) and (u) are hidden.

Claims

1. A computer-implemented method of generating personalized user models, the method comprising:

receiving automatic speech recognition (ASR) output of a plurality of speech interactions with a user;

receiving an ASR transcription error model characterizing how ASR transcription errors are made;

generating guesses of a true transcription and a user model via an expectation maximization (EM) algorithm based on the error model and the respective ASR output; and

generating a personalized user model based on the guesses.

2. The computer-implemented method of claim 1, wherein generating the guesses of the true transcription and a user model further comprises, iteratively:

alternating between generating a guess of the true transcription and of the user model; and

a current guess of one type is used to generate a next guess of the other type.

3. The computer-implemented method of claim 2, wherein iteratively generating the guesses of the true transcription and the user model until a threshold is met.

4. The computer-implemented method of claim 1, wherein the EM algorithm estimates conditional probabilities of hidden variables.

5. The computer-implemented method of claim 1, wherein generating the guesses is further based on a set of manual transcriptions of speech interactions with the user.

6. The computer-implemented method of claim 5, wherein the set of manual transcriptions is less numerous than the ASR output.

7. The method of claim 1, the method further comprising:

casting the speech interactions as a dynamical Bayesian network with four variables;

(s), (u), (r), and (m); and

encoding relationships between the (s), (u), (r), and (m) as conditional probability tables.

8. The computer-implemented method of claim 7, wherein at each dialog turn (r) and (m) are known and (s) and (u) are hidden.

9. The computer-implemented method of claim 1, the method further comprising generating a personalized ASR model based on the personalized user model.

10. The computer-implemented method of claim 1, the method further comprising recognizing additional speech from the user based on the personalized speech model.

11. The computer-implemented method of claim 10, the method further comprising iteratively improving the personalized speech model based on the additional speech.

12. The computer-implemented method of claim 1, wherein generating the personalized user model is further based on a previously generated personalized user model.

13. The computer-implemented method of claim 12, wherein the previously generated personalized user model is a template.

14. The computer-implemented method of claim 12, wherein the previously generated personalized user model is from a similar user.

15. The computer-implemented method of claim 1, wherein the ASR output is unlabeled.

16. A system for recognizing speech using personalized speech models, the system comprising:

a processor;

a module configured to control the processor to receive automatic speech recognition (ASR) output of a plurality of speech interactions with a user;

a module configured to control the processor to receive an ASR transcription error model characterizing how ASR transcription errors are made;

a module configured to control the processor to generate guesses of a true transcription and a user model via an expectation maximization (EM) algorithm based on the error model and the respective ASR output; and

a module configured to control the processor to generate a personalized user model based on the guesses.

17. The system of claim 16, wherein the ASR output is unlabeled.

18. The system of claim 16, wherein the module configured to control the processor to generate the guesses of the true transcription and the user model further comprises a module configured to iteratively perform the following steps:

alternating between generating a guess of the true transcription and of the user model; and

a current guess of one type is used to generate a next guess of the other type.

19. A computer-readable storage medium storing a computer program having instructions for controlling a processor to generate personalized user models, the instructions comprising:

receiving a user model personalized for a specific user generated by steps comprising;

receiving automatic speech recognition (ASR) output of a plurality of speech interactions with the specific user;

receiving an ASR transcription error model characterizing how ASR transcription errors are made;

generating guesses of a true transcription and a user model via an expectation maximization (EM) algorithm based on the error model and the respective ASR output;

generating a personalized user model for the specific user based on the guesses; and

building a personalized dialog system for the specific user based on the received personalized user model.

20. The computer-readable storage medium of claim 19, wherein the ASR output is unlabeled.