Speech recognition system and method

• US 4,394,538 A
• Filed: 03/04/1981
• Issued: 07/19/1983
• Est. Priority Date: 03/04/1981
• Status: Expired due to Term

• Current Assignee: Siemens Corporate Research & Support, Inc. (Siemens AG)
• Original Assignee: Threshold Technology, Inc.
• Inventors: Deng, Peter, White, George M., Warren, Ben, Parfitt, Rick
• Primary Examiner(s): KEMENY, EMANUEL
• Application Number:
  US06/240,273
• Time in Patent Office: 867 Days
• Field of Search: 179/1 SD, 179/1 SM, 179/1 SC, 340/146.3 WD, 340/146.3 AQ, 364/513
• US Class Current: 704/252
• CPC Class Codes: G10L 15/00 Speech recognition G10L17/00 takes precedence
• CPC Class Codes: G10L 15/063 Training

Abstract

An apparatus for receiving spoken input training words and for subsequently recognizing a spoken input command word. For each training word, and for the command word, there is stored a time dependent array and a cumulative feature non-time-dependent vector. The command word array is compared, member-by-member, with each training word array. Also, the command word vector is compared, feature-by-feature, with each training word vector. The results of both comparisons are used for selecting the training word which most closely matches the command word.

Claims

1. Apparatus for receiving spoken input training words and for subsequently recognizing a spoken input command word, comprising:

feature extraction means for generating feature signals indicative of the presence or absence of features in a spoken input word;

means for storing, as a time-dependent reference array, the presence or absence status of feature signals during each training word;

means for storing, as a non-time-dependent reference vector, an indication of the cumulative presence of feature signals during each training word;

means for storing, as a time-dependent feature array, the status of feature signals during the command word;

means for storing, as a non-time-dependent feature vector, an indication of the cumulative presence of feature signals during the command word;

means for comparing member-by-member, the command word feature array with each training word reference array to obtain an array comparison figure for each training word;

means for comparing, feature-by-feature, the command word feature vector with each training word reference vector to obtain a vector comparison figure for each training word; and

means responsive to both the array comparison figure and the vector comparison figure for each training word for selecting the training word which most closely matches the command word.

2. Apparatus as defined by claim 1 wherein said means for selecting said training word is responsive to a weighted sum of said array comparison figure and said vector comparison figure.

3. Apparatus as defined by claim 1 wherein said non-time-dependent reference vectors and feature vectors are representative of the normalized cumulative fractions of the spoken word duration during which individual feature components are present.

4. Apparatus as defined by claim 2 wherein said non-time-dependent reference vectors and feature vectors are representative of the normalized cumulative fractions of the spoken word duration during which individual feature components are present.

5. Apparatus as defined by claim 1 wherein said vector comparison figure SV_m for each training word m is determined as being ##EQU2## where n is the number of features in the vector, VR_mi are the individual vector components for the training word m, and VF_i are the individual vector components for the command word.

6. Apparatus as defined by claim 2 wherein said vector comparison figure SV_m for each training word m is determined as being ##EQU3## where n is the number of features in the vector, VR_mi are the individual vector components for the training word m, and VF_i are the individual vector components for the command word.

7. Apparatus as defined by claim 4 wherein said vector comparison figure SV_m for each training word m is determined as being ##EQU4## where n is the number of features in the vector, VR_mi are the individual vector components for the training word m, and VF_i are the individual vector components for the command word.

8. Apparatus for receiving spoken input training words and for subsequently recognizing a spoken input command word and generating an occurrence indication indicative of said command word, said occurrence indication being suitable for control of a companion system, comprising:

feature extraction means for generating feature signals indicative of the presence or absence of features in a spoken input word;

means for storing, as a time-dependent reference array, the presence or absence status of feature signals during each training word;

means for storing, as a non-time-dependent reference vector, an indication of the cumulative presence of feature signals during each training word;

means for storing, as a time-dependent feature array, the status of feature signals during the command word;

means for storing, as a non-time-dependent feature vector, an indication of the cumulative presence of feature signals during the command word;

means for comparing member-by-member, the command word feature array with each training word reference array to obtain an array comparison figure for each training word;

means for comparing, feature-by-feature, the command word feature vector with each training word reference vector to obtain a vector comparison figure for each training word; and

means responsive to both the array comparison figure and the vector comparison figure for each training word for selecting the training word which most closely matches the command word and for generating an occurrence indication that represents the selected training word.

9. Apparatus as defined by claim 8 wherein said means for selecting said training word is responsive to a weighted sum of said array comparison figure and said vector comparison figure.

10. Apparatus as defined by claim 8 wherein said non-time-dependent reference vectors and feature vectors are representative of the normalized cumulative fractions of the spoken word duration during which individual feature components are present.

11. Apparatus as defined by claim 9 wherein said non-time-dependent reference vectors and feature vectors are representative of the normalized cumulative fractions of the spoken word duration during which individual feature components are present.

12. Apparatus as defined by claim 8 wherein said vector comparison figure SV_m for each training word m is determined as being ##EQU5## where n is the number of features in the vector, VR_mi are the individual vector components for the training word m, and VF_i are the individual vector components for the command word.

13. Apparatus as defined by claim 9 wherein said vector comparison figure SV_m for each training word m is determined as being ##EQU6## where n is the number of features in the vector, VR_mi are the individual vector components for the training word m, and VF_i are the individual vector components for the command word.

14. Apparatus as defined by claim 11 wherein said vector comparison figure SV_m for each training word m is determined as being ##EQU7## where n is the number of features in the vector, VR_mi are the individual vector components for the training word m, and VF_i are the individual vector components for the command word.

15. Apparatus as defined by claim 9 wherein said occurrence indication is generated only when the combination of said array comparison figure and said vector comparison figure meets a predetermined standard.

16. Apparatus as defined by claim 10 wherein said occurrence indication is generated only when the combination of said array comparison figure and said vector comparison figure meets a predetermined standard.

17. Apparatus as defined by claim 11 wherein said occurrence indication is generated only when the combination of said array comparison figure and said vector comparison figure meets a predetermined standard.

18. A method for receiving spoken input training words and for subsequently recognizng a spoken input command word, comprising the steps of:

generating feature signals indicative of the presence or absence of features in a spoken input word;

storing, as a time-dependent reference array, the presence or absence status of feature signals during each training word;

storing, as a non-time-dependent reference vector, an indication of the cumulative presence of feature signals during each training word;

storing, as a time-dependent feature array, the status of feature signals during the command word;

storing, as a non-time-dependent feature vector, an indication of the cumulative presence of feature signals during the command word;

comparing member-by-member, the command word feature array with each training word reference array to obtain an array comparison figure for each training word;

comparing, feature-by-feature, the command word feature vector with each training word reference vector to obtain a vector comparison figure for each training word; and

selecting, as a function of both the array comparison figure and the vector comparison figure, the training word which most closely matches the command word.

19. The method as defined by claim 18 wherein said training word is selected as a function of a weighted sum of said array comparison figure and said vector comparison figure.

20. The method as defined by claim 18 wherein said non-time-dependent reference vectors and feature vectors are representative of the normalized cumulative fractions of the spoken word duration during which individual feature components are present.

21. The method as defined by claim 19 wherein said non-time-dependent reference vectors and feature vectors are representative of the normalized cumulative fractions of the spoken word duration during which individual feature components are present.

22. The method as defined by claim 18 wherein said vector comparison figure SV_m for each training word m is determined as being ##EQU8## where n is the number of features in the vector, VR_mi are the individual vector components for the training word m, and VF_i are the individual vector components for the command word.

23. The method as defined by claim 19 wherein said vector comparison figure SV_m for each training word m is determined as being ##EQU9## where n is the number of features in the vector, VR_mi are the individual vector components for the training word m, and VF_i are the individual vector components for the command word.

24. The method as defined by claim 21 wherein said vector comparison figure SV_m for each training word m is determined as being ##EQU10## where n is the number of features in the vector, VR_mi are the individual vector components for the training word m, and VF_i re the individual vector components for the command word.