System and method of using meta-data in speech processing

• US 20050096908A1
• Filed: 10/29/2004
• Published: 05/05/2005
• Est. Priority Date: 10/30/2003
• Status: Active Grant

• Current Assignee: Nuance Communications, Inc. (Microsoft Corporation)
• Original Assignee: AT&T Corporation (AT&T, Inc.)
• Inventors: Bacchiani, Michiel A.U., Sproat, Richard William, Roark, Brian E., Maskey, Sameer Raj
• Granted Patent: US 7,996,224 B2
• Time in Patent Office: Days
• US Class Current: 704/257
• CPC Class Codes: G10L 15/183 using context dependencies, e.g. language models
• CPC Class Codes: G10L 15/193 Formal grammars, e.g. finite state automata, context free grammars or word networks
• CPC Class Codes: G10L 15/197 Probabilistic grammars, e.g. word n-grams
• CPC Class Codes: G10L 15/22 Procedures used during a speech recognition process, e.g. man-machine dialogue
• CPC Class Codes: G10L 2015/228 of application context

Abstract

Systems and methods relate to generating a language model for use in, for example, a spoken dialog system or some other application. The method comprises building a class-based language model, generating at least one sequence network and replacing class labels in the class-based language model with the at least one sequence network. In this manner, placeholders or tokens associated with classes can be inserted into the models at training time and word/phone networks can be built based on meta-data information at test time. Finally, the placeholder token can be replaced with the word/phone networks at run time to improve recognition of difficult words such as proper names.

Claims

1. A method for generating a language model, the method comprising:

building a class-based language model;

generating at least one sequence network; and

replacing class labels in the class-based language model with the at least one sequence network.

2. The method of claim 1, wherein building the class-based language model further comprises replacing sequences of tokens from a training corpus for the class-based language model with the class labels.

3. The method of claim 2, wherein the class labels relate to at least one of a mailbox name and a caller name.

4. The method of claim 2, wherein the class labels are treated as words and building the class-based language model further comprises:

building a trigram model; and

encoding the trigram model as a weighted finite-state automaton.

5. The method of claim 4, wherein transitions in the weighted finite-state automaton are replaced by sequences of words that are members of that class.

6. The method of claim 1, wherein generating at least one sequence network further comprises:

estimating for each class of the class-based language model a probability of different realizations of words in each class; and

replacing a first weighted sequence network with the different realizations of words in each class to generate a second weighted sequence network.

7. The method of claim 6, wherein the words associated with the first weighted sequence network and the second weighted sequence network are names.

8. The method of claim 1, wherein the sequence network is a name sequence network.

9. The method of claim 1, wherein replacing class labels in the class-based language model with the at least one sequence network further comprises combining the weights of the at least one sequence network.

10. The method of claim 9, wherein combining the weights of the at least one sequence network is performed with a composition of finite-state transducers.

11. The method of claim 1, wherein each class label is a name class label that has a special phone symbol in a lexicon.

12. The method of claim 11, wherein replacing class labels in the class-based language model with the at least one sequence network further comprises:

combining the class-based language model with the lexicon;

optimizing the combined class-based language model and lexicon; and

replacing each transition in the class-based language model having a name class label as the output label with the optimized, combined class-based language model and lexicon for that name class.

13. The method of claim 12, further comprising combining weights in the replacement of each transition.

14. The method of claim 1, wherein building a class-based language model is performed at training time, generating at least one sequence network is performed at test time and replacing class labels in the class-based language model with the at least one sequence network is performed at run-time.

15. A speech recognition module using a language model, the language model generating by a method comprising:

building a class-based language model;

generating at least one sequence network; and

replacing class labels in the class-based language model with the at least one sequence network.

16. The speech recognition module of claim 15, wherein building the class-based language model further comprises replacing sequences of tokens from a training corpus for the class-based language model with the class labels.

17. The speech recognition module of claim 16, wherein the class labels are treated as words and building the class-based language model further comprises:

building a trigram model; and

encoding the trigram model as a weighted finite-state automaton.

18. The speech recognition module of claim 17, wherein transitions in the weighted finite-state automaton are replaced by sequences of words that are members of that class.

19. The speech recognition module of claim 15, wherein the sequence network is a name sequence network.

20. The speech recognition module of claim 15, wherein replacing class labels in the class-based language model with the at least one sequence network further comprises combining the weights of the at least one sequence network.

21. The speech recognition module of claim 15, wherein building a class-based language model is performed at training time, generating at least one sequence network is performed at test time and replacing class labels in the class-based language model with the at least one sequence network is performed at run-time.

22. A computer-readable medium that stores instructions for controlling a computing device to generate a language model, the instructions comprising the steps:

building a class-based language model;

generating at least one sequence network; and

replacing class labels in the class-based language model with the at least one sequence network.

23. The computer-readable medium of claim 22, wherein building the class-based language model further comprises replacing sequences of tokens from a training corpus for the class-based language model with the class labels.

24. The computer-readable medium of claim 22, wherein the class labels relate to at least one of a mailbox name and a caller name.

25. The computer-readable medium of claim 23, wherein the class labels are treated as words and building the class-based language model further comprises:

building a trigram model; and

encoding the trigram model as a weighted finite-state automaton.

26. The computer-readable medium of claim 25, wherein transitions in the weighted finite-state automaton are replaced by sequences of words that are members of that class.

27. The computer-readable medium of claim 22, wherein generating at least one sequence network further comprises:

estimating for each class of the class-based language model a probability of different realizations of words in each class; and

replacing a first weighted sequence network with the different realizations of words in each class to generate a second weighted sequence network.

28. The computer-readable medium of claim 27, wherein the words associated with the first weighted sequence network and the second weighted sequence network are names.

29. The computer-readable medium of claim 22, wherein the sequence network is a name sequence network.

30. The computer-readable medium of claim 22, wherein building a class-based language model is performed at training time, generating at least one sequence network is performed at test time and replacing class labels in the class-based language model with the at least one sequence network is performed at run-time.