Method of determining model-specific factors for pattern recognition, in particular for speech patterns

• US 6,456,969 B1
• Filed: 08/10/1999
• Issued: 09/24/2002
• Est. Priority Date: 12/12/1997
• Status: Expired due to Term

• Current Assignee: Nuance Communications, Inc. (Microsoft Corporation)
• Original Assignee: US Philips Corporation (Koninklijke Philips N.V.)
• Inventors: Beyerlein, Peter
• Primary Examiner(s): {haeck over (S)}mits, Ta̅livaldis Ivars
• Assistant Examiner(s): Azad, Abul K.
• Application Number:
  US09/367,214
• Time in Patent Office: 1,141 Days
• Field of Search: 704/243, 704/244, 704/245, 704/232, 704/233, 704/234, 704/240, 382/190
• US Class Current: 704/234
• CPC Class Codes: G06F 17/17 Function evaluation by approximation methods, e.g. inter- or extrapolation, smoothing, least mean square method G06F17/18 takes precedence ; interpolation for numerical control G05B19/18
• CPC Class Codes: G10L 15/142 Hidden Markov Models [HMMs]
• CPC Class Codes: G10L 15/183 using context dependencies, e.g. language models

Abstract

A method for recognizing a pattern that comprises a set of physical stimuli, said method comprising the steps of:

providing a set of training observations and through applying a plurality of association models ascertaining various measuring values p_j(k|x), j=1 . . . M, that each pertain to assigning a particular training observation to one or more associated pattern classes;

setting up a log/linear association distribution by combining all association models of the plurality according to respective weight factors, and joining thereto a normalization quantity to produce a compound association distribution;

optimizing said weight factors for thereby minimizing a detected error rate of the actual assigning to said compound distribution;

recognizing target observations representing a target pattern with the help of said compound distribution.

Claims

1. A method for modelling an association distribution for a plurality of patterns, said method comprising:

receiving a plurality of association models indicating various measuring values p_j(k|x), j=1 . . . M;

combining said plurality of association models in accordance with a set of weight factors to produce a log/linear association distribution;

joining a normalization quantity to said log/linear association distribution to produce a compound association distribution; and

optimizing said set of weight factors to minimize a detected error rate of an actual assigning to said compound association distribution, said optimizing of said set of weight factors being effected in a least squares method between an actual discriminant function and an ideal discriminant function, said actual discriminant function resulting from said compound association distribution, and said ideal discriminant function expressed on a basis of an error rate as smoothed through representing said error rate as a second degree curve in an interval (−

B,A); and

expressing a first weight factor Λ

of said set of weight factors in a closed expression Λ

=Q<sup>−

1</sup> P, wherein

said first weight factor Λ

is normalized through a constraining Σ

λ

_j=1,

said Q is an autocorrelation matrix of a set of discriminant functions of said plurality of association models as extended by an addition of a first normalization item, and

said P is an correlation vector between said error rate, said actual discriminant function and said ideal discriminant function as extended by an addition of a second normalization item.

2. The method of claim 1, further comprising:

receiving a set of training observations,

wherein each association model of said plurality of association models pertains to assigning a particular training observation of said set of training observations to at least one associated pattern class.

3. The method of claim 1, wherein

each association model of said plurality of association models is a first probability model, and

said compound association distribution is a second probability model for associating.

4. The method of claim 3, wherein

said first probability model is at least one model from a set of models including a language model, an acoustic model, a mllr adaption model, a unigram model, a distance-1 bigram model, a pentaphone model, and a wsj-model, and

said second probability model is at least one model from a set of models.

5. The method of claim 1, further comprising:

recognizing a set of target observations representing a target pattern with a help of said compound association distribution.

6. A system for modelling an association distribution for a plurality of patterns, said system comprising:

a storage means for storing a plurality of association models indicating various measuring values p_j(k|x), j=1 . . . M;

a first processing means for combining said plurality of association models in accordance with a set of weight factors to produce a log/linear association distribution;

a second processing means for joining a normalization quantity to said log/linear association distribution to produce a compound association distribution;

a third processing means for optimizing said set of weight factors to minimize a detected error rate of an actual assigning to said compound association distribution, said optimizing of said set of weight factors being effected in a least squares system between an actual discriminant function and an ideal discriminant function, said actual discriminant function resulting from said compound association distribution, and said ideal discriminant function expressed on a basis of an error rate as smoothed through representing said error rate as a second degree curve in an interval (−

B,A); and

a fourth processing means for expressing a first weight factor Λ

of said set of weight factors in a closed expression Λ

=Q<sup>−

1</sup>P, wherein

said first weight factor Λ

is normalized through a constraining Σ

λ

_j=1,

said Q is an autocorrelation matrix of a set of discriminant functions of said plurality of association models as extended by an addition of a first normalization item, and

said P is an correlation vector between said error rate, said actual discriminant function and said ideal discriminant function as extended by an addition of a second normalization item.

7. The system of claim 6, further comprising:

a pickup means for receiving a set of training observations,

where each association model of said plurality of association models pertains to assigning a particular training observation of said set of training observations to at least one associated pattern class.

8. The system of claim 6, wherein

each association model of said plurality of association models is a first probability model, and

said compound association distribution is a second probability model for associating.

9. The system of claim 8, wherein

said first probability model is at least one model from a set of models including a language model, an acoustic model, a mllr adaption model, a unigram model, a distance-1 bigram model, a pentaphone model, and a wsj-model, and

said second probability model is at least one model from a set of models.

10. The system of claim 6, further comprising:

a recognizing means for recognizing a set of target observations representing a target pattern with a help of said compound association distribution.