Word recognizer

• US 4,400,828 A
• Filed: 03/27/1981
• Issued: 08/23/1983
• Est. Priority Date: 03/27/1981
• Status: Expired due to Term

• Current Assignee: Bell Telephone Laboratories, Inc. (Nokia Corporation)
• Original Assignee: Bell Telephone Laboratories, Inc. (Nokia Corporation)
• Inventors: Rabiner, Lawrence R., Wilpon, Jay G., Pirz, Frank C.
• Primary Examiner(s): Boudreau, Leo H.
• Application Number:
  US06/248,547
• Time in Patent Office: 879 Days
• Field of Search: 340/146.3 WD, 340/146.3 AQ, 340/146.3 Q, 340/146.3 M, 340/146.3 A, 179/1 SB, 179/1 SD, 179/1 SA, 382/1, 382/16, 382/30, 382/33, 382/34, 382/36, 343/5 MM, 358/105-108
• US Class Current: 704/238
• CPC Class Codes: G10L 15/00 Speech recognition G10L17/00 takes precedence
• CPC Class Codes: G10L 15/12 using dynamic programming techniques, e.g. dynamic time warping [DTW]

Abstract

An input word is recognized as one of a set of reference words. A set of word distance signals representative of the correspondence of the input word to the reference words is generated. A set of weighted word distance signals is also generated. Responsive to the word distance signals and the weighted word distance signals, the reference word that most closely corresponds to the input word is selected.

Claims

1. Apparatus for recognizing an input word as one of a set of reference words comprising

means for storing a plurality of reference word feature templates representative of said reference words;

means responsive to said input word for generating an input word feature template;

means responsive to said reference word feature templates and said input word feature template for identifying said input word as one of said reference words, characterized in that

said input word identifying means (100) comprises

means (105,200) responsive to said reference word feature templates and said input word feature template for generating a set of word distance signals;

said reference word feature templates and said input word feature template each have a plurality of frames;

said means for generating a set of word distance signals comprises means responsive to said reference word feature templates and said input word feature template for generating a set of frame distance signals representative of the similarity between frames of said reference word feature templates and said input word feature template, and means for combining said frame distance signals to produce said word distance signals;

means (105, 200,

     500) responsive to said reference word feature templates and said input word feature template for generating a set of weighted word distance signals; and

means (300,

     400) responsive to said word distance signals and said weighted word distance signals for selecting the reference word which most corresponds to said input word.

2. Apparatus as in claim 1 further characterized in that

said reference words for which feature templates are stored belong to a predetermined set of equivalence classes, each class being representative of reference words which are within a prescribed degree of similarity;

and

said means for generating a set of weighted word distance signals comprises

means for storing a plurality of weight templates representative of the expected similarity between frames of reference word feature templates for reference words which belong to the same equivalence class;

means responsive to said frame distance signals and said weight templates for generating weighted frame distance signals; and

means for combining said weighted frame distance signals to produce said weighted word distance signals.

3. Apparatus as in claim 2 further characterized in that

said means for selecting the reference word which most corresponds to said input word comprises

means responsive to said word distance signals for selecting the equivalence class of the reference word representative of the word distance signal which most corresponds to said input word,

means responsive to said weighted word distance signals for identifying the reference word of the selected equivalence class representative of the weighted word distance signal which most corresponds to said input word,

whereby said identified reference word designates a candidate for recognition of the input word.

4. Apparatus as in claims 1 or 2 further characterized in that

said input word feature template and said reference word feature templates are representative of acoustic characteristics of spoken words.

5. Apparatus as in claims 2 or 3 further characterized in that said equivalence classes are representative of reference words which are acoustically similar, and

said frames are representative of discrete periods of time.

6. Apparatus as in claim 5 further characterized in that said reference word feature templates and said input word feature template each have the same number of time frames.

7. A method for recognizing an input word as one of a set of reference words comprising the steps of

storing a plurality of reference word feature templates representative of said reference words;

generating an input word feature template responsive to said input word;

identifying said input word as one of said reference words responsive to said reference word feature templates and said input word feature template, characterized in that

said input word identifying step comprises

generating a set of word distance signals responsive to said reference word feature templates and said input word feature template;

said reference word feature templates and said input word feature template each have a plurality of frames;

said step for generating a set of word distance signals comprises generating a set of frame distance signals representative of the similarity between frames of said reference word feature templates and said input word feature template responsive to said reference word feature templates and said input word feature template, and combining said frame distance signals to produce said word distance signals;

generating a set of weighted word distance signals responsive to said reference word feature templates and said input word feature template; and

selecting the reference word which most corresponds to said input word responsive to said word distance signals and said weighted word distance signals.

8. A method for recognizing an input word as one of a set of reference words according to claim 7 further characterized in that

said reference words for which feature templates are stored belong to a predetermined set of equivalence classes, each class being representative of reference words which are within a prescribed degree of similarity;

and

said step for generating a set of weighted word distance signals comprises

storing a plurality of weight templates representative of the expected similarity between frames of reference word feature templates for reference words which belong to the same equivalence class;

generating weighted frame distance signals responsive to said frame distance signals and said weight templates; and

combining said weighted frame distance signals to obtain said weighted word distance signals.

9. A method for recognizing an input word as one of a set of reference words according to claim 8 further characterized in that

said step for selecting the reference word which most corresponds to said input word comprises

selecting the equivalence class of the reference word representative of the word distance signal which most corresponds to said input word responsive to said word distance signals;

identifying the reference word of the selected equivalence class representative of the weighted word distance signal which most corresponds to said input word responsive to said weighted word distance signals,

whereby said identified reference word designates a candidate for recognition of the input word.

10. A method for recognizing an input word as one of a set of reference words according to claims 7 or 8 further characterized in that

said input word feature template and said reference word feature templates are representative of acoustic characteristics of spoken words.

11. A method for recognizing an input word as one of a set of reference words according to claims 8 or 9 further characterized in that

said equivalence classes are representative of reference words which are acoustically similar, and

said frames are representative of discrete periods of time.

12. A method for recognizing an input word as one of a set of reference words according to claim 11 further characterized in that said reference word feature templates and said input word feature template each have the same number of time frames.

13. Apparatus for recognizing a spoken input word as one of a set of reference words comprising

means for storing a plurality of reference word feature templates representative of acoustic characteristics of said reference words;

means responsive to said input word for generating an input word feature template representative of acoustic characteristics of said input word;

means responsive to said reference word feature templates and said input word feature template for identifying said input word as one of said reference words;

said input word identifying means including means responsive to said reference word feature templates and said input word feature templates for generating a set of word distance signals;

said reference word feature templates and said input word feature template each have a plurality of frames;

said reference words for which feature templates are stored belong to a predetermined set of equivalence classes, each class being representative of reference words which are within a prescribed degree of similarity;

said means for generating a set of word distance signals comprises means responsive to said reference word feature templates and said input word feature template for generating a set of frame distance signals representative of the similarity between frames of said reference word feature templates and said input word feature template, and means for summing said frame distance signals to product said word distance signals;

means responsive to said reference word feature templates and said input word feature template for generating a set of weighted word distance signals;

said means for generating a set of weighted word distance signals comprises means for storing a plurality of weight templates representative of the expected similarity between frames of reference word feature templates for reference words which belong to the same equivalence class, means responsive to said frame distance signals and said weight templates for generating weighted frame distance signals, and means for summing said weighted frame distance signals to produce said weighted word distance signals; and

means responsive to said word distance signals and said weighted word distance signals for selecting the reference word which most corresponds to said input word.

14. Apparatus as in claim 13 wherein said means for selecting the reference word which most corresponds to said input word comprises:

means responsive to said word distance signals for selecting the equivalence class of the reference word representative of the minimum word distance signal;

means responsive to said weighted word distance signals for identifying the reference word of the selected equivalence class representative of the minimum weighted word distance signal;

whereby said identified reference word designates a candidate for recognition of the input word.

15. A method for recognizing a spoken input word as one of a set of reference words comprising

storing a plurality of reference word feature templates representative of acoustic characteristics of said reference words;

generating an input word feature template representative of acoustic characteristics of said input word responsive to said input word;

identifying said input word as one of said reference words responsive to said reference word feature templates and said input word feature template;

said input word identifying step comprises generating a set of word distance signals responsive to said reference word feature templates and said input word feature template;

said reference word feature templates and said input word feature template each have a plurality of frames;

said reference words for which feature templates are stored belong to a predetermined set of equivalence classes, each class being representative of reference words which are within a prescribed degree of similarity;

said step for generating a set of word distance signals comprises generating a set of frame distance signals representative of the similarity between frames of said reference word feature templates and said input word feature template responsive to said reference word feature templates and said input word feature template, and summing said frame distance signals to produce said word distance signals;

generating a set of weighted word distance signals responsive to said reference word feature templates and said input word feature template;

said step for generating a set of weighted word distance signals comprises storing a plurality of weight templates representative of the expected similarity between frames of reference word feature templates for reference words which belong to the same equivalence class, generating weighted frame distance signals responsive to said frame distance signals and said weight templates, and summing said weighted frame distance signals to produce said weighted word distance signals; and

selecting the reference word which most corresponds to said input word responsive to said word distance signals and said weighted word distance signals.

16. A method for recognizing a spoken input word as one of a set of reference words according to claim 15 wherein said step for selecting the reference word which most corresponds to said input word comprises:

selecting the equivalence class of the reference word representative of the minimum word distance signal responsive to said word distance signals;

identifying the reference word of the selected equivalence class representative of the minimum weighted word distance signal responsive to said weighted word distance signals;

whereby said identified reference word designates a candidate for recognition of the input word.