Technique for adaptation of hidden markov models for speech recognition

• US 6,151,574 A
• Filed: 09/08/1998
• Issued: 11/21/2000
• Est. Priority Date: 12/05/1997
• Status: Expired due to Fees

• Current Assignee: Lucent Technologies, Inc. (Nokia Corporation)
• Original Assignee: Lucent Technologies, Inc. (Nokia Corporation)
• Inventors: Shinoda, Koichi, Lee, Chin-Hui
• Primary Examiner(s): Zele, Krista
• Assistant Examiner(s): Opsasnick, Michael N.
• Application Number:
  US09/149,782
• Time in Patent Office: 805 Days
• Field of Search: 704/243, 704/244, 704/245, 704/256, 382/178-185
• US Class Current: 704/256
• CPC Class Codes: G10L 15/065 Adaptation
• CPC Class Codes: G10L 15/144 Training of HMMs

Abstract

A speech recognition system learns characteristics of speech by a user during a learning phase to improve its performance. Adaptation data derived from the user'"'"'s speech and its recognized result is collected during the learning phase. Parameters characterizing hidden Markov Models (HMMs) used in the system for speech recognition are modified based on the adaptation data. To that end, a hierarchical structure is defined in an HMM parameter space. This structure may assume the form of a tree structure having multiple layers, each of which includes one or more nodes. Each node on each layer is connected to at least one node on another layer. The nodes on the lowest layer of the tree structure are referred to as "leaf nodes." Each node in the tree structure represents a subset of the HMM parameters, and is associated with a probability measure which is derived from the adaptation data. In particular, each leaf node represents a different one of the HMM parameters, which is derivable from the probability measure associated with the leaf node. This probability measure is a function of the probability measures which are associated with the nodes connected to the leaf node, and which represent "hierarchical priors" to such a probability measure.

Claims

1. Apparatus for recognizing selected speech based on acoustic models, the apparatus comprising:

a processor responsive to selected data representing a sample of the selected speech for modifying the acoustic models; and

a mechanism for defining a hierarchical structure which includes a plurality of levels, each level having one or more nodes thereon, each level in the structure arranged higher than every other level having a smaller number of nodes thereon than the other level, each node being associated with a probability measure which is determined based on at least the selected data, each node on each level having two or more nodes thereon being connected to a node on a second level higher than the level having two or more nodes thereon, the probability measure associated with each node on each level having two or more nodes thereon being a function of at least the probability measure associated with the node connected thereto on the second level, the acoustic models being modified based on at least the probability measure associated with each node on a selected level.

2. The apparatus of claim 1 wherein the acoustic models are in accordance with hidden Markov models (HMMs).

3. The apparatus of claim 2 wherein the HMMs include continuous density HMMs (CDHMMs).

4. The apparatus of claim 1 wherein the probability measure associated with each node is determined based also on second data representing a sample of at least second speech.

5. The apparatus of claim 4 wherein the probability measure associated with each node is derived from a difference between the sample of the selected speech and that of the at least second speech, the difference being a function of the selected data and second data.

6. The apparatus of claim 1 wherein the acoustic models are modified in response also to a recognized version of the sample of the speech.

7. The apparatus of claim 1 wherein the acoustic models are characterized by a plurality of parameters, each node representing a set containing one or more of the parameters.

8. The apparatus of claim 7 wherein the selected level is arranged lowest of all of the levels, each node on the selected level representing a set containing a different one of the plurality of parameters.

9. The apparatus of claim 8 wherein the probability measure associated with each node on the selected level is used to modify the parameter in the set represented by the node.

10. The apparatus of claim 8 wherein the set represented by each node on a level other than the selected level includes the sets represented by the nodes on the selected level which are connected to the node.

11. A system for recognizing selected speech based on acoustic models, which are characterized by a plurality of parameters, the system comprising:

a device for providing selected data representing a sample of the selected speech;

a processor for defining a structure which includes a plurality of levels, each level including one or more nodes, each node being associated with a respective probability measure, which is derived from at least the selected data;

a mechanism for identifying at least one sequence of nodes from different levels; and

an adaptor for modifying at least one of the parameters based on at least the probability measure associated with a selected node in the sequence, the probability measure associated with the selected node being a function of the probability measure associated with every other node in the sequence.

12. The system of claim 11 wherein the levels in the structure are arranged in order in a direction, a level arranged further in the direction including a larger number of nodes, the selected node being from the level arranged furthest in the direction of all of the levels.

13. The system of claim 11 wherein the probability measure associated with the selected node is also a function of at least one weight for scaling the probability measure associated with a second node in the sequence, the weight varying with a position of the level where the second node is from, relative to positions of other levels in the structure.

14. The system of claim 13 wherein the weight also varies with an amount of the data.

15. The system of claim 11 wherein the acoustic models are in accordance with HMMs.

16. The system of claim 11 wherein the HMMs include CDHMMs.

17. The system of claim 11 wherein the respective probability measure is derived also from second data representing a sample of at least second speech.

18. The apparatus of claim 17 wherein the respective probability measure is determined based on a difference between the sample of the selected speech and that of the at least second speech, the difference being a function of the selected data and second data.

19. The system of claim 11 wherein the acoustic models are modified based also on a recognized version of the sample of the speech.

20. A method for use in an apparatus for recognizing selected speech based on acoustic models, the method comprising:

modifying the acoustic models in response to selected data representing a sample of the selected speech; and

defining a hierarchical structure which includes a plurality of levels, each level having one or more nodes thereon, each level in the structure arranged higher than every other level having a smaller number of nodes thereon than the other level, each node being associated with a probability measure which is determined based on at least the selected data, each node on each level having two or more nodes thereon being connected to a node on a second level higher than the level having two or more nodes thereon, the probability measure associated with each node on each level having two or more nodes thereon being a function of at least the probability measure associated with the node connected thereto on the second level, the acoustic models being modified based on at least the probability measure associated with each node on a selected level.

21. The method of claim 20 wherein the acoustic models are in accordance with HMMs.

22. The method of claim 21 wherein the HMMs include continuous density CDHMMs.

23. The method of claim 20 wherein the probability measures are determined based also on second data representing a sample of at least second speech.

24. The method of claim 23 wherein the probability measure associated with each node is derived from a difference between the sample of the selected speech and that of the at least second speech, the difference being a function of the selected data and second data.

25. The method of claim 20 wherein the acoustic models are modified in response also to a recognized version of the sample of the speech.

26. The method of claim 20 wherein the acoustic models are characterized by a plurality of parameters, each node representing a set containing one or more of the parameters.

27. The method of claim 26 wherein the selected level is arranged lowest of all of the levels, each node on the selected level representing a set containing a different one of the plurality of parameters.

28. The method of claim 27 wherein the probability measure associated with each node on the selected level is used to modify the parameter in the set represented by the node.

29. The method of claim 27 wherein the set represented by each node on a level other than the selected level includes the sets represented by the nodes on the selected level which are connected to the node.

30. A method for use in a system for recognizing selected speech based on acoustic models, which are characterized by a plurality of parameters, the method comprising:

providing selected data representing a sample of the selected speech;

defining a structure which includes a plurality of levels, each level including one or more nodes, each node being associated with a respective probability measure, which is derived from at least the selected data;

identifying at least one sequence of nodes from different levels; and

modifying at least one of the parameters based on at least the probability measure associated with a selected node in the sequence, the probability measure associated with the selected node being a function of the probability measure associated with every other node in the sequence.

31. The method of claim 30 wherein the levels in the structure are arranged in order in a direction, a level arranged further in the direction including a larger number of nodes, the selected node being from the level arranged furthest in the direction of all of the levels.

32. The method of claim 30 wherein the probability measure associated with the selected node is also a function of at least one weight for scaling the probability measure associated with a second node in the sequence, the weight varying with a position of the level where the second node is from, relative to positions of other levels in the structure.

33. The method of claim 32 wherein the weight also varies with an amount of the data.

34. The method of claim 30 wherein the acoustic models are in accordance with HMMs.

35. The method of claim 30 wherein the HMMs include CDHMMs.

36. The method of claim 30 wherein the respective probability measure is derived also from second data representing a sample of at least second speech.

37. The method of claim 36 wherein the respective probability measure is determined based on a difference between the sample of the selected speech and that of the at least second speech, the difference being a function of the selected data and second data.

38. The method of claim 30 wherein the acoustic models are modified based also on a recognized version of the sample of the speech.