Apparatus and method for normalizing and categorizing linear prediction code vectors using Bayesian categorization technique

• US 5,704,004 A
• Filed: 01/21/1997
• Issued: 12/30/1997
• Est. Priority Date: 12/01/1993
• Status: Expired due to Term

• Current Assignee: Industrial Technology Research Institute
• Original Assignee: Industrial Technology Research Institute
• Inventors: Pengwu, Chung-Mou, Chen, Cheng-Der, Li, Tze Fen, Sun, Chung-Yu
• Primary Examiner(s): MacDonald, Allen R.
• Assistant Examiner(s): CHAWAN, VIJAY B
• Application Number:
  US08/786,551
• Time in Patent Office: 343 Days
• Field of Search: 395/2.52, 395/2.28, 395/2.43, 395/2.45
• US Class Current: 704/243
• CPC Class Codes: G10L 15/02 Feature extraction for speech recognition; Selection of recognition unit
• CPC Class Codes: G10L 2015/027 Syllables being the recognition units
• CPC Class Codes: G10L 2015/0635 updating or merging of old and new templates; Mean values; Weighting
• CPC Class Codes: G10L 25/12 the extracted parameters being prediction coefficients
• CPC Class Codes: G10L 25/24 the extracted parameters being the cepstrum

Abstract

The present invention discloses a pattern matching system applicable for syllable recognition which includes a dictionary means for storing a plurality of standard patterns each representing a standard syllable by at least a syllable feature. The pattern matching system further includes a converting means for converting an input pattern representing an unknown syllable into a categorizing pattern for representing the unknown syllable in the syllable features used for representing the standard syllables. The pattern matching system further includes a Bayesian categorizing means for matching the standard pattern representing the standard syllable and the categorizing pattern representing the unknown syllable for computing a Bayesian mis-categorization risk for each of the standard syllables, the Bayesian categorization means further including a comparing and identification means for selecting a standard syllable which has the least mis-categorization risk as an identified syllable for the input unknown syllable.

Claims

1. A pattern matching system provided for performing a sequence of single syllables recognition comprising:

a dictionary means for storing a plurality of standard patterns wherein each of said standard patterns representing a single standard syllable by a set of feature vectors C(1), C(2), C(3), . . . , and C(M) and M being a positive integer;

a converting means for converting an input pattern representing single unknown syllable into a categorizing pattern for representing said single unknown syllable in a set of categorizing vectors X where X={x(1), x(2),x(3), . . . ,x(k)} where k representing a positive integer; and

a Bayesian-decision-rule categorizing means for computing a conditional normal density function ƒ

(x| Ci) for each of said feature vectors Ci, wherein said function ƒ

(x| Ci) having a normal distribution and said x(1), x(2), x(3), . . . and x(k) are stochastically independent; and

said Bayesian-decision-rule categorizing means further employing functional parameters of said normal distribution for said normal density function ƒ

(x| Ci) to apply a Bayesian decision rule to deterministically identify said single unknown syllable with one of said standard single syllables.

2. The pattern matching system of claim 1 wherein:

said Bayesian-decision-rule categorizing means further computing said conditional normal density function ƒ

(x| Ci) as;

##EQU10## where i=1,2,3. . . , M and M is number of syllables to be recognized;

said Bayesian-decision-rule categorizing means further computing logarithmic values of ƒ

(x1, . . . ,xk| Ci) for comparing values of L (C₁) where;

##EQU11## for deterministically identifying a category Ci which has the least L(Ci) as said standard syllable for identifying with said single unknown syllable.

3. The pattern matching system of claim 2 further comprises:

a linear predictive coding (LPC) means for converting each of said standard patterns into a LPC cepstra, and for converting said categorizing patterns into a categorizing LPC cepstra vector; and

a compression means for compressing said LPC cepstra vector into a compressed standard pattern represented by said of feature vectors C(1), C(2), C(3), . . . ,C(M), for storing in said dictionary means, and for compressing said categorizing LPC cepstra vectors into a compressed categorizing patterns represented by said set of categorizing vectors X where X={(x(1),x(2),x(3), . . . ,x(k)} for storing in said dictionary means.

4. The pattern matching system of claim 2 further comprises:

an user interface means to allow an user of said pattern matching system to provide input data and commands for controlling the operation of said matching system.

5. The pattern matching system of claim 4 further comprises:

a receiver for receiving an incoming syllable utterance in the form of wavefunctions; and

an analog to digital conversion means for converting said wavefunctions to a plurality of digital data representing said wavefunctions.

6. A pattern matching system provided for performing a sequence of single syllables recognition comprising:

a receiver for receiving an incoming syllable utterance of either a standard syllable or a single unknown syllable for syllable recognition in the form of wavefunctions;

an analog to digital conversion means for converting said wavefunctions to a plurality of digital data representing said wave functions;

a linear predictive coding (LPC) means for converting said digital data representing said input wavefunction into a LPC cepstra vector;

a compression means for compressing said each LPC cepstra vector into a compressed cepstra vector wherein said unknown syllable is represented by a set of categorizing vectors X where X={x(1),x(2),x(3), . . . ,x(k)};

a dictionary means for storing a plurality of standard compressed cepstra vectors each representing a standard single syllable by a set of feature vectors C(1), C(2), C(3), . . . , and C(M);

a Bayesian-decision-rule categorizing means for computing a conditional normal density function ƒ

(x| Ci) for each of said feature vectors Ci, assuming that said function ƒ

(x| Ci) having a normal distribution and said x(1), x(2), x(3), . . . and x(k) are stochastically independent;

,

said Bayesian-decision-rule categorizing means further employing functional parameters of said normal distribution for said normal density function ƒ

(x| Ci) to apply a Bayesian decision rule to deterministically identify said single unknown syllable with one of said standard syllables; and

an user interface means to low an user of said pattern matching system to provide input data and commands for controlling the operation of said matching system.

7. The pattern matching system of claim 6 wherein:

said compression means compressing said LPC cepstra vectors represented by Yk where Yk={y(k)₁, y(k)₂,y(k)₃, . . . ,y (k)_p } and K=1,2,3 . . . , n, by deleting a stable portion of said vectors with a difference of two of said consecutive LPC cepstra vectors be denoted as;

##EQU12## for k=2, 3, 4, . . . , n, and by deleting one of said LPC cepstra vectors when said D(K) is below a pre-designated threshold value.

9. The pattern matching system of claim 6 wherein:

said compression means compressing said LPC cepstra vectors represented by Yk where Yk={y(k)₁, y(k)₂,y(k)₃, . . . ,y (k)_p } and K=1,2,3 . . . , n, by normalizing each of said LPC cepstra vectors by applying a total sum of absolute values of said LPC cepstra vector S wherein;

##EQU14## and by dividing a syllable into a M sections wherein M is a positive integer and said LPC cepstra vectors for each section are normalized to an absolute value of S/M.

8. The pattern matching system of daim 7 wherein:

said compression means compressing said LPC cepstra vectors represented by Y'"'"'k where Y'"'"'k={y'"'"'(k)₁, y'"'"'(k)₂, y'"'"'(k)₃, . . . , y'"'"'(k)_p } and K=1,2,3 . . . , m, according to a sum S1 of absolute differences of two consecutive LPC cepstra vectors wherein;

##EQU13## and said LPC cepstra vectors are divided into a plurality sections with an average value of said LPC cepstra vectors in each of section characterized by a sum-difference-feature of said section.

10. A method for matching and categorizing an input pattern of waveform applicable for syllable recognition comprising the steps of:

.

(a) storing in a dictionary means a plurality of standard patterns wherein each of said standard patterns representing a standard single syllable by a set of feature vectors C(1), C(2), C(3), . . . , and C(M) and M being a positive integer;

(b) converting said input pattern of waveform representing an unknown single syllable into a categorizing pattern for representing said unknown syllable by a set of categorizing vectors X where X={x(1), x(2), x(3), . . . ,x(k)} and k being a positive integer;

(c) utilizing a Bayesian-decision-rule categorizing means for computing a conditional normal density function ƒ

(x| Ci) wherein said function ƒ

(x| Ci) having a normal distribution and said x(1), x(2), x(3), . . . and x(k) are stochastically independent;

, and

(d) employing functional parameters of said normal distribution for said normal density function ƒ

(x| Ci) to apply a Bayesian decision rule to identify said input unknown single syllable with one of said standard single syllables.

11. The method for matching and categorizing an input pattern of waveform as recited in claim 10 wherein:

said step (a) further includes a step of converting each of said standard patterns into a standard LPC cepstra vector; and

said step of converting said input pattern of waveform representing an unknown single syllable into a categorizing pattern is a step of converting said input pattern of waveform into a categorizing LPC cepstra vector.

12. The method for matching and categorizing an input pattern of waveform as recited in claim 11 wherein:

said step (a) further includes a step of compressing each of said standard cepstra vectors into a set of standard compressed cepstra vectors represented by said of feature vectors C(1), C(2), C(3), . . . , and C(M), for storing in said dictionary means; and

said step (c) further includes a step of compressing said categorizing LPC cepstra vector into a compressed categorizing LPC cepstra vector represent by said set of categorizing vectors X where X={x(1), x(2),x(3), . . . ,x(k)}.

13. The method for matching and categorizing an input pattern of waveform as recited in claim 12 further comprises a step of:

(e) providing an user interface means to allow an user of said method for matching and categorizing said input pattern of waveform to input data and commands for controlling the operation of said method.

14. The method for matching and categorizing an input pattern of waveform as recited in claim 13 further comprises the steps of:

(a1) providing a receiver for receiving said input pattern of waveform; and

(a2) utilizing an analog to digital conversion means for converting said input pattern of waveform to a plurality of digital data representing said input pattern of waveform wherein said steps (a1) and (a2) are performed prior to the performance of said step (a).