Minimizing empirical error training and adaptation of statistical language models and context free grammar in automatic speech recognition

• US 20080255844A1
• Filed: 04/10/2007
• Published: 10/16/2008
• Est. Priority Date: 04/10/2007
• Status: Active Grant

• Current Assignee: Microsoft Technology Licensing LLC (Microsoft Corporation)
• Original Assignee: Microsoft Corporation
• Inventors: Wu, Jian
• Granted Patent: US 7,925,505 B2
• Time in Patent Office: Days
• US Class Current: 704/255
• 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 2015/0631 Creating reference templates; Clustering

Abstract

Architecture for minimizing an empirical error rate by discriminative adaptation of a statistical language model in a dictation and/or dialog application. The architecture allows assignment of an improved weighting value to each term or phrase to reduce empirical error. Empirical errors are minimized whether a user provides correction results or not based on criteria for discriminatively adapting the user language model (LM)/context-free grammar (CFG) to the target. Moreover, algorithms are provided for the training and adaptation processes of LM/CFG parameters for criteria optimization.

Claims

1. A computer-implemented system that facilitates speech recognition, comprising:

a recognition component for generating a recognized result based on an input phrase, the recognition component using a statistical language model (SLM) having an empirical error rate;

an interaction component for interacting with the recognized result to create a corrected result; and

an adaptation component for discriminatively adapting the SLM to the corrected result based on criteria.

2. The system of claim 1, wherein the criteria for minimizing the empirical error rate is an objective function defined by,

$L=\frac{1}{N}$

∑

i = 1 N 

l 

( X i , W c , W r ) .

where L is averaged over an entire training corpus, N is a number of training samples, l is a loss function, X_i is an i-th sample of training, W_c is a correct word transcription based on user correction activity and/or pre-labeled training corpus, and W_r is a competitive word string received from incorrect recognition results of a recognition engine and against which comparison of the correct word transcription is made.

3. The system of claim 2, wherein the objective function is minimized by a sequential gradient-based solution.

4. The system of claim 3, wherein parameter values used for minimizing the objective function are employed in at least one of a dialog application or a dictation application.

5. The system of claim 1, wherein the interaction component facilitates creation of the corrected result by user correction of the recognized result.

6. The system of claim 1, wherein the adaptation component computes an approximate distance between a probability distribution of a correct word transcription and a probability distribution of an incorrect word transcription.

7. The system of claim 1, wherein the adaptation component facilitates discriminative adaptation and training of context-free grammars (CFG) to optimize the criteria.

8. The system of claim 1, wherein the adaptation component utilizes an acoustical model to generate a distance between a probability distribution of a correct word transcription and a probability distribution of an incorrect word transcription if there are insufficient training samples.

9. The system of claim 1, wherein the adaptation component optimizes n-gram parameters when acoustical data is available.

10. The system of claim 1, wherein the corrected result is weighted against the recognized result to create a score, the score fed back to the recognition component for recognition processing of additional input phrases.

11. A computer-implemented method of processing speech, comprising:

processing input terms into first-time recognized results using a recognizer and statistical language model;

generating user-corrected results based on the first-time recognized results;

generating new language model scores based on the user-corrected results and the first-time recognized results; and

inputting the new language model scores to a recognizer to process additional input terms.

12. The method of claim 11, further comprising weighting the first-time recognized results based on the new scores to reduce empirical error in the language model.

13. The method of claim 11, further comprising discriminatively training CFG parameters in a dialog application.

14. The method of claim 11, further comprising discriminatively adapting CFG parameters in a dialog application.

15. The method of claim 11, further comprising discriminatively adapting the statistical language model in a dictation application.

16. The method of claim 11, further comprising discriminatively adapting CFG parameters independent of acoustic data availability.

17. The method of claim 16, further comprising employing information of at least one of a lexicon model or an acoustical model to approximate a difference between the first-time recognition results and the user-corrected results.

18. The method of claim 11, further comprising optimizing CFG weights to minimize empirical error rate when no acoustical data is available.

19. The method of claim 11, further comprising batch processing the additional input terms based on the minimized empirical error rate.

20. A computer-implemented system, comprising:

computer-implemented means for processing input terms into first-time recognized results using a statistical language model;

computer-implemented means for generating user-corrected results based on the first-time recognized results;

computer-implemented means for generating new language model scores based on the user-corrected results and the first-time recognized results; and

computer-implemented means for inputting the new language model scores back to a recognizer to process additional input terms.