Work Files Saved Searches
   My Account                                                  Search:   Quick/Number   Boolean   Advanced       Help   


 The Delphion Integrated View

  Buy Now:   Buy PDF- 15pp  PDF  |   File History  |   Other choices   
  Tools:  Citation Link  |  Add to Work File:    
  View:  Expand Details   |  INPADOC   |  Jump to: 
 
 Email this to a friend  Email this to a friend 
       
Title: US6044344: Constrained corrective training for continuous parameter system
[ Derwent Title ]


Country: US United States of America

View Images High
Resolution

 Low
 Resolution

 
15 pages

 
Inventor: Kanevsky, Dimitri; Ossining, NY

Assignee: International Business Machines Corporation, Armonk, NY
other patents from INTERNATIONAL BUSINESS MACHINES CORPORATION (280070) (approx. 44,393)
 News, Profiles, Stocks and More about this company

Published / Filed: 2000-03-28 / 1997-01-03

Application Number: US1997000778800

IPC Code: Advanced: G01L 9/00;
IPC-7: G01L 9/00;

U.S. Class: Current: 704/256.7; 704/244;
Original: 704/256; 704/244;

Field of Search: 704/231,255,256,251 714/244

Priority Number:
1997-01-03  US1997000778800

Abstract:     A method is provided for training a statistical pattern recognition decoder on new data while preserving its accuracy of old, previously learned data. Previously learned data are represented as constrained equations that define a constrained domain (T) in a space of statistical parameters (K) of the decoder. Some part of a previously learned data is represented as a feasible point on the constrained domain. A training procedure is reformulated as optimization of objective functions over the constrained domain. Finally, the constrained optimization functions are solved. This training method ensures that previously learned data is preserved during iterative training steps. While an exemplary speech recognition decoder is discussed, the inventive method is also suited to other pattern recognition problems such as, for example, handwriting recognition, image recognition, machine translation, or natural language processing.

Attorney, Agent or Firm: Whitham, Curtis & Whitham ; Otterstedt, Paul J. ;

Primary / Asst. Examiners: Hudspeth, David R.; Zintel, Harold

Maintenance Status: E1 Expired  Check current status

INPADOC Legal Status: Show legal status actions

Family: None

First Claim:
Show all 6 claims
I claim:     1. A computer implemented method of corrective training for a speech recognition system that is modeled by a continuous parameter system, comprising the steps of:
  • choosing a space of continuous parameter systems in which every point in said space gives rise to a separate continuous parameter system;
  • transforming a textual script and a sound message that was generated by a human speaker from said textual script into electrical signal representations;
  • producing from said electrical representations a system of constrained equations that define a domain T in said space of continuous parameter systems satisfying following:
    • said continuous parameter system comprises a point A, that satisfies said system of constrained equations;
    • said speech recognition system that is modeled by a continuous parameter system that corresponds to any point in said domain T decodes with not worse accuracy than said speech recognition system that is modeled by a continuous parameter system that corresponds to the point A;
  • representing a training procedure as a constrained optimization problem of a likelihood score function defined over said domain T; and
  • finding a new continuous parameter system in T that solves said constrained optimization problem and constructing a new speech recognition system that is modeled by said new continuous parameter system, wherein said new continuous parameter system is used to recognize sound used in said new speech recognition system for recognition of sound messages or for repeating said method of corrective training with new speech data,
  • wherein said step of finding a new continuous parameter system in T that solves said constrained optimization problem comprises the steps:
    • taking an iterative training step to find a point B, where B is in the space of said continuous parameter system, not necessarily contained in the constrained domain T;
    • taking an interval I to find a new data point belonging to the constrained domain T and lying on an interval connecting point A and point B, such that any point on the interval I has a higher likelihood score than the likelihood score corresponding to A; and
    • repeating said step of taking an interval I to find a new data point until said new data point converges to a local maximum in the constrained domain T.


Background / Summary: Show background / summary

Drawing Descriptions: Show drawing descriptions

Description: Show description

Forward References: Show 7 U.S. patent(s) that reference this one

       
U.S. References: Go to Result Set: All U.S. references   |  Forward references (7)   |   Backward references (0)   |   Citation Link

       
Foreign References:
Buy
PDF
Publication Date IPC Code Assignee   Title
Get PDF - 11pp EP0786761 1997-07  G01L 5/06 AT & T CORP Method of speech recognition using decoded state sequences having constrained state likelihoods 


Other References:
  • Baum et al "A maximization technique occurring in the statistical analysis of probabilistic functions of markov chains" 1970, The annals of mathematics statistics, vol. 41, No. 1, pp. 164-171. Cited by 13 patents
  • Kanevsky "A generalization of the Baum algorithim to functions on non-linear manifolds" May 1995, 1995 conference on acoustics, speech, and signal processing, vol. 1, pp. 473-476.
  • Salgado et al "Optimal power flow solutions using the gradient projection method" Nov. 1990, IEE proceedings, pp. 424-428. (5 pages) Cited by 2 patents
  • Gopalakrishnan et al. An Inequality for Rational Functions with Applications to Some Statistical Estimation Problems, IEEE TRabsactions on Informatioon Theory, vol. 37 No. 1, Jan. 1991.
  • Deller "Discrete-time processing of speech signals" Prentice-Hall pp. 692,705-714,738, 1987.
  • Bahl et al "A maximum likelihood approach to continuous speech recognition" IEEE pp. 179-190, Mar. 1983. (12 pages) Cited by 42 patents


  • Inquire Regarding Licensing

    Powered by Verity


    Plaques from Patent Awards      Gallery of Obscure PatentsNominate this for the Gallery...

    Thomson Reuters Copyright © 1997-2014 Thomson Reuters 
    Subscriptions  |  Web Seminars  |  Privacy  |  Terms & Conditions  |  Site Map  |  Contact Us  |  Help