System and method for using meta-data dependent language modeling for automatic speech recognition
View Patent ↗Disclosed are systems and methods for providing a spoken dialog system using meta-data to build language models to improve speech processing. Meta-data is generally defined as data outside received speech; for example, meta-data may be a customer profile having a name, address and purchase history of a caller to a spoken dialog system. The method comprises building tree clusters from meta-data and estimating a language model using the built tree clusters. The language model may be used by various modules in the spoken dialog system, such as the automatic speech recognition module and/or the dialog management module. Building the tree clusters from the meta-data may involve generating projections from the meta-data and further may comprise computing counts as a result of unigram tree clustering and then building both unigram trees and higher-order trees from the meta-data as well as computing node distances within the built trees that are used for estimating the language model.
1. A method of generating a language model using non-speech metadata, the method comprising:
applying a tree growing algorithm to highly fragmented non-speech metadata associated with a caller to a spoken dialog system, wherein the highly fragmented non-speech metadata does not describe physical characteristics of the caller;
identifying projections based on the highly fragmented non-speech metadata and leaf nodes in which a history appears as a result of applying the tree growing algorithm to yield identified projections; and
estimating, via a processor of a computing device, a conditional, metadata dependent language model based on the identified projections and speech data from the caller.
2. The method of claim 1 , wherein identifying projections further comprises using divisive likelihood-based techniques to identify non-speech metadata projections.
3. The method of claim 2 , wherein identifying projections further comprises using tree-based clustering to identify the non-speech metadata projections.
4. The method of claim 1 , wherein the estimated conditional, metadata dependent language model is estimated by merging non-speech metadata conditional counts that are weighted based on a given non-speech metadata value.
5. The method of claim 1 , wherein identifying the projections is performed at training time.
6. The method of claim 5 , wherein generating the estimated conditional metadata dependent language model is performed at test time.
7. The method of claim 1 , wherein identifying the projections further comprises estimating a unigram model to generate a unigram clustering output.
8. The method of claim 7 , further comprising building a higher order tree from the unigram clustering output.
9. The method of claim 8 , wherein building the higher order tree further comprises:
generating history nodes for the higher order tree; and
splitting leaves on the higher order tree via domain splits.
10. The method of claim 9 , wherein generating history nodes further comprises:
partitioning count distributions allowing only history dependent splits; and
terminating the splitting based on a threshold.
11. The method of claim 8 , further comprising computing a distance between a root node of a unigram tree and at least one history node of a higher order tree.
12. The method of claim 11 , wherein generating the estimated conditional, metadata dependent language model further comprises weighing the identified projections using the distance.
13. The method of claim 7 , wherein estimating the unigram model further comprises:
generating a unigram tree by splitting a leaf with a largest likelihood gain; and
evaluating a likelihood gain of new leaf created by the splits.
14. The method of claim 13 , wherein splitting the leaf with the largest likelihood gain is iterative.
15. The method of claim 14 , further comprising terminating the iterative splitting process if one the likelihood gain of a best leaf split falls below a threshold, wherein the unigram tree defines N leaf sets.
16. The method of claim 1 , wherein generating the estimated conditional, metadata dependent language model further comprises estimating history dependent distributions for histories in training data.
17. A method of performing automatic speech recognition, the method comprising:
receiving speech;
performing automatic speech recognition on the speech using a language model generated by steps comprising:
applying a tree growing algorithm to highly fragmented non-speech metadata associated with a caller to a spoken dialog system, wherein the highly fragmented non-speech metadata does not describe physical characteristics of the caller;
identifying projections based on the highly fragmented non-speech metadata and leaf nodes in which a history appears as a result of applying the tree growing algorithm to yield identified projections; and
estimating a conditional, metadata dependent language model based on the identified projections and speech data from the caller.
18. The method of claim 17 , wherein identifying projections further comprises using divisive likelihood-based techniques to identify non-speech metadata projections.
19. A system for automatic speech recognition, the system comprising:
a processor; and
a first module configured to control the processor to recognize received speech using a language model generated by steps comprising:
applying a tree growing algorithm to highly fragmented non-speech metadata associated with a caller to a spoken dialog system, wherein the highly fragmented non-speech metadata does not describe physical characteristics of the caller;
identifying projections based on the highly fragmented non-speech metadata and leaf nodes in which a history appears as a result of applying the tree growing algorithm to yield identified projections; and
estimating a conditional, metadata dependent language model based on the identified projections and speech data from the caller.
20. The system of claim 19 , wherein the conditional, metadata dependent language model is estimated by merging non-speech metadata conditional counts that are weighted based on a given non-speech metadata value.