Method of optimizing a composite speech recognition expert

• US 5,280,563 A
• Filed: 12/20/1991
• Issued: 01/18/1994
• Est. Priority Date: 12/20/1991
• Status: Expired due to Term

• Current Assignee: Nuance Communications, Inc. (Microsoft Corporation)
• Original Assignee: Kurzweil Applied Intelligence, Inc. (Intel Corporation)
• Inventors: Ganong, William F.
• Primary Examiner(s): Fleming, Michael R.
• Assistant Examiner(s): Doerrler, Michelle
• Application Number:
  US07/812,581
• Time in Patent Office: 760 Days
• Field of Search: 381/29-46, 395/2
• US Class Current: 704/200
• CPC Class Codes: G10L 15/18 using natural language modelling
• CPC Class Codes: G10L 15/197 Probabilistic grammars, e.g. word n-grams

Abstract

In a continuous speech recognizer which includes at least, one acoustic expert and one linguistic expert which generate respective scores, a method is disclosed for adjusting the relative weighting to be applied to those scores employing training data utilizing the words to be recognized in multiple word phrases. Multiple word test phrases are applied to the acoustic expert to determine, for each phrase, plural multi-word hypotheses each having corresponding cumulative scores. The linguistic expert generates corresponding cumulative linguistic scores. An objective function is calculated for each test phrase having a value which is variable as a function of the difference between the combined score of any correct hypothesis and that of the most easily confused incorrect hypothesis. The objective function values are cumulated and a gradient descent procedure is used to adjust the relative weighting of the acoustic and linguistic scores in obtaining a combined score.

Claims

1. The method of adjusting the relative weighting coefficients to be applied to respective scores generated in a continuous speech recognizer which incorporates plural experts including at least an acoustic expert and a linguistic expert, the acoustic and linguistic experts generating respective scores for each work postulated from the system'"'"'s vocabulary;

said method comprising;

obtaining continuous speech training data utilizing the words to be recognized in multiple word phrases;

submitting said multiple word phrases to at least one acoustic expert thereby to determine, for each such phrase, plural multi-word hypotheses each having a corresponding cumulative acoustic score denoting likelihood of match;

submitting said hypotheses to at least one linguistic expert thereby to determine for each hypothesis a corresponding cumulative linguistic score denoting likelihood of match;

for each hypothesis, combining said acoustic and linguistic cumulative scores using weighting coefficients initially arbitrarily selected thereby to obtain a respective combined hypothesis score;

for each submitted multiple work phrase, determining if any hypothesis is correct and, if a correct hypothesis exists,

identifying the hypothesis which has the best combined score of those hypotheses which are not the correct hypothesis

computing an objective function having a value which is variable as a function of a difference calculated between the combined score of the correct hypothesis and the combined score of said identified incorrect hypothesis;

cumulating the objective function values over all phrases which produced at least one correct hypothesis; and

calculating a vector direction of adjustment of said relative weighting coefficients which improves the cumulated objective function values; and

making an adjustment of said relative weighting coefficients in that direction.

2. The method as set forth in claim 1 wherein the objective function corresponds to a summation over all words and over the plural experts of the weighted scores for each word.

3. The method as set forth in claim 2 wherein the cumulation of objective function values corresponds to a summation over all test phrases of a difference calculated between the objective function value for the correct hypothesis and the objective function value for the hypothesis which has the best combined score of those hypotheses which are not the correct hypothesis.

4. The method as set forth in claim 3 wherein a sigmoidal function is applied to each difference before summation.

5. The method as set forth in claim 4 wherein said sigmoidal function is defined essentially as:

##EQU10## where x is the respective difference.

6. The method of adjusting the relative weighting coefficients to be applied to respective scores generated in a continuous speech recognizer which incorporates plural experts including at least an acoustic expert and a linguistic expert, the acoustic and linguistic experts generating respective scores for each word postulated from the system'"'"'s vocabulary;

said method comprising;

obtaining continuous speech training data utilizing the words to be recognized in multiple word phrases;

a) submitting said multiple work phrases to at least one acoustic expert thereby to determine, for each such phrase, plural multi-work hypotheses each having a corresponding cumulative acoustic score denoting likelihood of match;

b) submitting said hypotheses to at least one linguistic expert thereby to determine for each hypothesis a corresponding cumulative linguistic score denoting likelihood of match

c) for each hypothesis, combining said acoustic and linguistic cumulative scores using weighting coefficients initially arbitrarily selected thereby to obtain a respective combined hypothesis score;

d) for each submitted phrase, determining if any hypothesis is correct and, if a correct hypothesis exists, computing an objective function having a value which is variable as a function of a difference calculated between the combined score of the correct hypothesis and the combined score of the hypothesis which has the best score of those hypotheses which are not the correct hypothesis;

e) cumulating the objective function values over all phrases which produced at least one correct hypothesis; and

f) calculating a vector direction of adjustment of said coefficients which improves the cumulated objective function values;

g) making an adjustment of said coefficients in said direction; and

repeating steps a)-g) at least once.

7. The method of adjusting the relative weighting coefficients to be applied to respective scores generated in a continuous speech recognizer which incorporates plural experts including at least an acoustic expert and a linguistic expert, the acoustic and linguistic experts generating respective scores for each work postulated from the system'"'"'s vocabulary;

said method comprising;

obtaining continuous speech training data utilizing the words to be recognized in multiple word phrases;

submitting said multiple word phrases to at least one acoustic expert thereby to determine, for each such phrase, plural multi-word hypotheses each having a corresponding cumulative acoustic score denoting likelihood of match;

submitting said hypotheses to at least one linguistic expert thereby to determine for each hypothesis a corresponding cumulative linguistic score denoting likelihood of match;

for each hypothesis, combining said acoustic and linguistic cumulative scores using weighting coefficients initially arbitrarily selected thereby to obtain a respective combined hypothesis score;

for each submitted multiple work phrase, determining if any hypothesis is correct and, if a correct hypothesis exists,

identifying the hypothesis which has the best combined score of those hypotheses which are not the correct hypothesis

between the combined score of the correct hypothesis and the combined score of said identified incorrect hypothesis;

adjusting said difference in accordance with a sigmoidal function ##EQU11## where x is the respective difference computing an objective function having a value which is variable as a function of the adjusted difference

calculating a difference

cumulating the objective function values over all phrases which produced at least one correct hypothesis; and

calculating a vector direction of adjustment of said relative weighting coefficients which improved the cumulated objective function values; and

making an adjustment of said relative weighting coefficients in that direction.