Language model data collection

Patent metadata

Field	Value
Publication number	US-9047868-B1
Application number	US-201213563648-A
Country	US
Kind code	B1
Filing date	Jul 31, 2012
Priority date	Jul 31, 2012
Publication date	Jun 2, 2015
Grant date	Jun 2, 2015

Abstract

A specific language model for speech recognition may be built. In some embodiments, the specific language model is associated with a user and built using a corpus of text obtained from a user computing device. In some embodiments, a sequence of words is constructed from the corpus of text. The sequence of words may be obfuscated, and the obfuscated sequence of words may be stored in the specific language model. A server or a user device may use the specific language model in conjunction with a general language model to perform speech recognition on an utterance made by the user.

First claim

The invention claimed is:

1. A system comprising:
an electronic data store configured to store data related to a user; and
a text processing device in communication with the electronic data store, the text processing device configured to:
obtain text data from a user computing device;
construct an n-gram based on the text data;
generate, for the n-gram, a hash value;
determine, for the n-gram, a first count indicating a number of times the n-gram appears within the text data;
add the first count to a second count, wherein the second count is associated with the hash value; and
store the hash value and the second count in the electronic data store; and
transmit the hash value and the second count to a server device over a network, wherein the server device is configured to:
receive the hash value and the second count;
determine a specific n-gram probability using the hash value and the second count;
obtain a general n-gram probability from a general language model;
determine a combined n-gram probability using the specific n-gram probability and the general n-gram probability; and
perform speech recognition using the combined n-gram probability.

2. The system of claim 1 , wherein the text processing device is configured to obtain text data from the user computing device by monitoring network traffic generated by the user computing device.

3. The system of claim 1 , wherein the text processing device is configured to obtain text data from data received directly from the user computing device.

4. The system of claim 1 , wherein a number of bits of the hash value is based on a privacy setting.

5. The system of claim 1 , wherein the n-gram comprises a trigram.

6. The system of claim 1 , wherein the text processing device is further configured to store metadata associated with the n-gram in the electronic data store, wherein metadata associated with the n-gram is one of a source of the n-gram, information about an entity associated with the n-gram, or a date and time when the text data was obtained.

7. A non-transitory computer-readable medium comprising a module configured to execute in one or more processors of a computing device, the module being further configured to:
receive text data, wherein the text data originated from a user computing device;
construct an n-gram based on the text data;
generate, for the n-gram, a hash value, wherein a number of bits of the hash value is based on a privacy setting;
determine language model information associated with the hash value;
store the hash value and the language model information in an electronic data store; and
transmit the hash value and the language model information to a server device.

8. The non-transitory computer-readable medium of claim 7 , wherein the language model information is a count.

9. The non-transitory computer-readable medium of claim 7 , wherein the language model information is a tri-gram probability.

10. The non-transitory computer-readable medium of claim 7 , wherein the server device is configured to store the hash value and the language model information in a language model.

11. The non-transitory computer-readable medium of claim 7 , wherein the module is further configured to monitor network traffic generated by the user computing device, and wherein the text data comprises network traffic generated by the user computing device.

12. The non-transitory computer-readable medium of claim 7 , wherein the network interface module is further configured to receive the text data directly from the user computing device.

13. The non-transitory computer readable medium of claim 7 , wherein the computing device comprises the user computing device.

14. The non-transitory computer-readable medium of claim 7 , wherein the language model building module is further configured to store metadata associated with the n-gram in the electronic data store, wherein the metadata associated with the n-gram comprises a source from which the n-gram is generated, an entity associated with the n-gram, or a date and time when the text data was generated.

15. The non-transitory computer readable medium of claim 7 , wherein the text data comprises at least one of an electronic message, a document, a music library, or a contact list.

16. A computer-implemented method comprising:
as implemented by a server device configured with specific computer-executable instructions,
obtaining a first n-gram probability for an n-gram from a first language model;
determining, by the server device executing a language model building module, a hash value for the n-gram;
obtaining a second n-gram probability from a second language model using the hash value;
determining a third n-gram probability using the first n-gram probability and the second n-gram probability; and
performing speech recognition using the third n-gram probability.

17. The computer-implemented method of claim 16 , wherein determining the third n-gram probability comprises using language model interpolation.

18. The computer-implemented method of claim 16 , wherein obtaining the first n-gram probability comprises determining a second hash value for the n-gram and obtaining the first n-gram probability from a language model using the second hash value.

19. The computer-implemented method of claim 18 , wherein the second hash value has a larger number of bits than the hash value.

20. The computer-implemented method of claim 16 , wherein the first language model is a general purpose language model and the second language model is a specific language model adapted to a user.

21. The computer-implemented method of claim 16 , wherein determining a third n-gram probability using the first n-gram probability and the second n-gram probability comprises performing a batch operation prior to performing speech recognition.

22. A system comprising:
a server device configured to:
obtain a first n-gram probability for an n-gram from a first language model;
determine a hash value for the n-gram;
obtain a second n-gram probability from a second language model using the hash value;
determine a third n-gram probability using the first n-gram probability and the second n-gram probability; and
perform speech recognition using the third n-gram probability.

23. The system of claim 22 , wherein the server device is configured to determine the third n-gram probability by using language model interpolation.

24. The system of claim 22 , wherein the server device is configured to obtain the first n-gram probability by determining a second hash value for the n-gram and obtaining the first n-gram probability from a language model using the second hash value.

25. The system of claim 24 , wherein the second hash value has a larger number of bits than the hash value.

26. The system of claim 22 , wherein the first language model is a general purpose language model and the second language model is a specific language model adapted to a user.

27. The system of claim 22 , wherein the server device is configured to determine a third n-gram probability using the first n-gram probability and the second n-gram probability by performing a batch operation prior to performing speech recognition.

28. A system comprising:
an electronic data store configured to store data related to a user; and
a text processing device in communication with the electronic data store, t

Assignees

• O'Neill Jeffrey C
• Thomas Ryan P
• Amazon Tech Inc

Inventors

• O'Neill Jeffrey C
• Thomas Ryan P

Classifications

• G10L15/197 Primary

  Probabilistic grammars, e.g. word n-grams · CPC title

• G10L15/30

  Distributed recognition, e.g. in client-server systems, for mobile phones or network applications · CPC title