Speech recognition apparatus and method

• US 4,319,085 A
• Filed: 04/08/1980
• Issued: 03/09/1982
• Est. Priority Date: 04/08/1980
• Status: Expired due to Term

• Current Assignee: Siemens Corporate Research & Support, Inc. (Siemens AG)
• Original Assignee: Threshold Technology, Inc.
• Inventors: Welch, John R., Herscher, Marvin B.
• Primary Examiner(s): Atkinson, Charles E.
• Assistant Examiner(s): KEMENY, EMANUEL
• Application Number:
  US06/138,646
• Time in Patent Office: 700 Days
• Field of Search: 179/1 SD, 179/1 SB, 179/1 VC, 340/146.3 WD, 340/146.3 SG, 340/148, 364/410, 364/519, 364/513
• US Class Current: 704/255
• CPC Class Codes: G10L 19/04 using predictive techniques
• CPC Class Codes: G10L 25/87 Detection of discrete points within a voice signal

Abstract

In this speech recognition system the array formed by a timewise sequence of speech signal feature vectors includes digital data at each time slot representing both presence/absence and consistency of occurence.

Claims

1. Apparatus which receives input spoken vocabulary words during a training phase of operation and subsequently recognizes received input spoken command words, comprising:

feature extraction means for processing received input words and generating digital feature signals dependent upon the features present in said input words;

means for forming, for each vocabulary word, a time dependent reference array having digital information at each array position representative of the presence or absence of a particular feature signal at a particular time slot during at least a predetermined fraction of a number of training utterances of said vocabulary word and also representative of the consistency of occurrence of said particular feature signal at said particular time slot during said number of training utterances of said vocabulary word;

means for forming a time dependent command word feature array having digital information at each array position representative of the presence or absence of a particular feature signal at a particular time slot during a command word candidate;

means for comparing, member-by-member, the command word feature array with the reference array for each vocabulary word; and

means for selecting the vocabulary word those reference array yields the highest correlation with said command word feature array.

2. Apparatus as defined by claim 1 wherein said reference array for each vocabulary word includes two binary bits at each array position, one of said bits representing the presence or absence of a particular feature at a particular time slot, and the other of said bits representing the consistency of presence or absence of said particular feature at said particular time slot.

3. Apparatus as defined by claim 2 wherein said feature array for each command word candidate has one binary bit at each array position representing the presence or absence of a particular feature at a particular time slot.

4. Apparatus as defined by claim 1 wherein said means for forming reference arrays and command word feature arrays include means for time-normalizing said reference arrays and said command word arrays.

5. Apparatus as defined by claim 2 wherein said means for forming reference arrays and command word feature arrays include means for time-normalizing said reference arrays and said command word arrays.

6. Apparatus as defined by claim 3 wherein said means for forming reference arrays and command word feature arrays include means for time-normalizing said reference arrays and said command word arrays.

7. Apparatus as defined by claim 2 wherein said means for performing said member-by-member comparison includes:

means for comparing each presence or absence indicative bit of the command word feature array with the corresponding presence or absence indicative bit of the reference array; and

means for assigning a first predetermined correlation weight to a match of feature presence, a second predetermined correlation weight to a match of feature absence and either a third or fourth predetermined correlation weight to a mismatch of bits;

the weight assigned to a mismatch having a value which depends upon the status of the consistency bit at the corresponding array position of the reference array.

8. Apparatus as defined by claim 5 wherein said means for performing said member-by-member comparison includes:

means for comparing each presence or absence indicative bit of the command word feature array with the corresponding presence or absence indicative bit of the reference array; and

means for assigning a first predetermined correlation weight to a match of feature presence, a second predetermined correlation weight to a match of feature absence and either a third or fourth predetermined correlation weight to a mismatch of bits;

the weight assigned to a mismatch having a value which depends upon the status of the consistency bit at the corresponding array position of the reference array.

9. Apparatus as defined by claim 6 wherein said means for performing said member-by-member comparison incudes:

means for comparing each presence or absence indicative bit of the command word feature array with the corresponding presence or absence indicative bit of the reference array; and

means for assigning a first predetermined correlation weight to a match of feature presence, a second predetermined correlation weight to a match of feature absence and either a third or fourth predetermined correlation weight to a mismatch of bits;

the weight assigned to a mismatch having a value which depends upon the status of the consistency bit at the corresponding array position of the reference array.

10. Apparatus as defined by claim 7 wherein a match of feature absence is assigned zero correlation weight, a match of feature presence is assigned a positive correlation weight, and a mismatch is assigned a negative correlation weight, the magnitude of said negative correlation weight being relatively small when said consistency bit indicates inconsistency and being relatively large when said consistency bit indicates consistency.

11. Apparatus as defined by claim 8 wherein a match of feature absence is assigned zero correlation weight, a match of feature presence is assigned a positive correlation weight, and a mismatch is assigned a negative correlation weight, the magnitude of said negative correlation weight being relatively small when said consistency bit indicates inconsistency and being relatively large when said consistency bit indicates consistency.

12. Apparatus as defined by claim 9 wherein a match of feature absence is assigned zero correlation weight, a match of feature presence is assigned a positive correlation weight, and a mismatch is assigned a negative correlation weight, the magnitude of said negative correlation weight being relatively small when said consistency bit indicates inconsistency and being relatively large when said consistency bit indicates consistency.

13. A method for receiving input spoken vocabulary words during a training phase of operation and subsequently recognizing input spoken command words, comprising the steps of:

generating digital feature signals dependent upon the features present in said received input words;

forming, for each vocabulary word, a time dependent reference array having digital information at each array position representative of the presence or absence of a particular feature signal at a particular time slot during at least a predetermined fraction of a number of training utterances of said vocabulary word and also representative of the consistency of occurrence of said particular feature signal at said particular time slot during said number of training utterances of said vocabulary word;

forming a time dependent command word feature array having digital information at each array position representative of the presence or absence of a particular feature at a particular time slot during a command word candidate;

comparing, member-by-member, the command word feature array with the reference array for each vocabulary word; and

selecting the vocabulary word whose reference array yields the highest correlation with said command word feature array.

14. The method as defined by claim 13 wherein said reference array for each vocabulary word includes two binary bits at each array position, one of said bits representing the presence or absence of a particular feature at a particular time slot, and the other of said bits representing the consistency of presence or absence of said particular feature at said particular time slot.

15. The method as defined by claim 14 wherein said feature array for each command word candidate has one binary bit at each array position representing the presence or absence of a particular feature at a particular time slot.

16. The method as defined by claim 13 wherein said steps of forming reference arrays and command word feature arrays include time-normalizing said reference arrays and said command word arrays.

17. The method as defined by claim 14 wherein said steps of forming reference arrays and command word feature arrays include time-normalizing said reference arrays and said command word arrays.

18. The method as defined by claim 15 wherein said steps of forming reference arrays and command word feature arrays include time-normalizing said reference arrays and said command word arrays.

19. The method as defined by claim 14 wherein said step of performing said member-by-member comparison includes:

comparing each presence or absence indicative bit of the command word feature array with the corresponding presence or absence indicative bit of the reference array; and

assigning a first predetermined correlation weight to a match of feature presence, a second predetermined correlation weight to a match of feature absence and either a third or fourth predetermined correlation weight to a mismatch of bits;

the weight assigned to a mismatch having a value which depends upon the status of the consistency bit at the corresponding array position of the reference array.

20. The methods as defined by claim 17 wherein said step of performing said member-by-member comparison includes:

comparing each presence or absence indicative bit of the command word feature array with the corresponding presence or absence indicative bit of the reference array; and

assigning a first predetermined correlation weight to a match of feature presence, a second predetermined correlation weight to a match of feature absence and either a third or fourth predetermined correlation weight to a mismatch of bits;

the weight assigned to a mismatch having a value which depends upon the status of the consistency bit at the corresponding array position of the reference array.

21. The method as defined by claim 18 wherein said step of performing said member-by-member comparison includes:

comparing each presence or absence indicative bit of the command word feature array with the corresponding presence or absence indicative bit of the reference array; and

assigning a first predetermined correlation weight to a match of feature presence, a second predetermined correlation weight to a match of feature absence and either a third or fourth predetermined correlation weight to a mismatch of bits;

the weight assigned to a mismatch having a value which depends upon the status of the consistency bit at the corresponding array position of the reference array.

22. The method as defined by claim 19 wherein a match of feature absence is assigned zero correlation weight, a match of feature presence is assigned a positive correlation weight, and a mismatch is assigned a negative correlation weight, the magnitude of said negative correlation weight being relatively small when said consistency bit indicates inconsistency and being relatively large when said consistency bit indicates consistency.

23. The method as defined by claim 20 wherein a match of feature absence is assigned zero correlation weight, a match of feature presence is assigned a positive correlation weight, and a mismatch is assigned a negative correlation weight, the magnitude of said negative correlation weight being relatively small when said consistency bit indicates inconsistency and being relatively large when said consistency bit indicates consistency.

24. The method as defined by claim 21 wherein a match of feature absence is assigned zero correlation weight, a match of feature presence is assigned a positive correlation weight, and a mismatch is assigned a negative correlation weight, the magnitude of said negative correlation weight being relatively small when said consistency bit indicates inconsistency and being relatively large when said consistency bit indicates consistency.