Voice language model adjustment based on user affinity

• US 7,590,536 B2
• Filed: 10/07/2005
• Issued: 09/15/2009
• Est. Priority Date: 10/07/2005
• Status: Expired due to Fees

• Current Assignee: Nuance Communications, Inc. (Microsoft Corporation)
• Original Assignee: Nuance Communications, Inc. (Microsoft Corporation)
• Inventors: Wallenfelt, Brian P., Bates, Cary L.
• Primary Examiner(s): Hudspeth; David R
• Assistant Examiner(s): GUERRA-ERAZO, EDGAR X
• Application Number:
  US11/246,831
• Publication Number:
  US 20070083374A1
• Time in Patent Office: 1,439 Days
• Field of Search: 704/257, 704/263, 704/266, 704/270.1, 704/255, 704/275, 704/277, 704/231, 704243-245
• US Class Current: 704/244
• CPC Class Codes: G10L 15/063 Training
• CPC Class Codes: G10L 15/183 using context dependencies, e.g. language models
• CPC Class Codes: G10L 15/197 Probabilistic grammars, e.g. word n-grams

Abstract

Methods, systems and computer readable medium for improving the accuracy of voice processing are provided. Embodiments of the present invention generally provide methods, systems and articles of manufacture for adjusting a language model within a voice recognition system. In one embodiment, changes are made to the language model by identifying a word-usage pattern that qualifies as an anomaly. In one embodiment, an anomaly occurs when the use of a given word (or phrase) differs from an expected probability for the word (or phrase), as predicted by a language model. Additionally, observed anomalies may be shared and applied by different users of the voice processing system, depending on an affinity in word-usage frequency between different users.

Claims

1. A computer-implemented method for increasing the accuracy of a voice recognition system, comprising:

initiating a voice processing session for a first user, wherein the voice processing session is configured to match a voice pattern derived from one or more words spoken by the first user with entries in a word-usage probability table, based on a probability of the one or more words being spoken by the first user;

during the voice processing session initiated for the first user, monitoring for an occurrence of a word-usage anomaly, wherein the word-usage anomaly occurs when an observed usage frequency for a given one or more words differs, by a predetermined magnitude, from an expected frequency recorded in the word-usage probability table for the given one or more words;

in response to the occurrence of the word usage anomaly, increasing the expected frequency for the given one or more words in the word-usage probability table, wherein the increased expected frequency for the given one or more words in the word-usage probability table is configured to expire, after a predetermined period of time;

transmitting an indication of the observed anomaly to a voice processing session initiated for a second user; and

monitoring, by a server, respective voice processing sessions for each of a plurality of users, including at least the first user and the second user, wherein the server is configured to;

(i) monitor the occurrence of word-usage anomalies for each of the voice processing sessions,

(ii) record anomalies observed by each of the voice processing sessions in a server anomaly table, and

(iii) broadcast an indication of a global anomaly to each of the voice processing sessions, wherein the global anomaly comprises an anomaly that is observed independently by at least two of the voice processing sessions.

2. The computer-implemented method of claim 1, wherein the probability of the one or more words being spoken by the first user are determined according to an n-gram model, and wherein the probabilities in the word-usage probability table comprise an n-gram probability regarding a sequence of n words.

3. The computer-implemented method of claim 1, further comprising:

determining, by the voice processing system initiated for the second user, a correlation score indicating how often anomalies observed by the first user were also observed by the second user; and

depending on the correlation score, adjusting the expected frequency for the word corresponding to the observed anomaly in a word-usage probability table initialized for the voice processing session of the second user.

4. The computer-implemented method of claim 1, wherein the word-usage probability table is initialized for the second user using at least one indication of a word-usage anomaly observed by the voice processing session of the first user.

5. The computer-implemented method of claim 1, wherein monitoring for an occurrence of a word-usage anomaly comprises monitoring corrections made to a word selected by the voice processing session to match with a given voice pattern that are entered manually by the user.

6. The computer-implemented method of claim 1, wherein the increased expected frequency for the given one or more words in the word-usage probability table is configured to diminish, over time, to return the increased expected frequency to the original expected frequency.