Start/end point detection for word recognition

• US 5,794,195 A
• Filed: 05/12/1997
• Issued: 08/11/1998
• Est. Priority Date: 06/28/1994
• Status: Expired due to Fees

• Current Assignee: Alcatel NV (Nokia Corporation)
• Original Assignee: Alcatel NV (Nokia Corporation)
• Inventors: Hormann, Thomas, Rozinaj, Gregor
• Primary Examiner(s): Hunter, Daniel S.
• Application Number:
  US08/854,472
• Time in Patent Office: 456 Days
• Field of Search: 704/233, 704/247, 704/248, 704/251, 704/252, 704/253, 704/226
• US Class Current: 704/253
• CPC Class Codes: G10L 15/05 Word boundary detection
• CPC Class Codes: G10L 2025/783 based on threshold decision
• CPC Class Codes: G10L 25/12 the extracted parameters being prediction coefficients
• CPC Class Codes: G10L 25/24 the extracted parameters being the cepstrum
• CPC Class Codes: G10L 25/87 Detection of discrete points within a voice signal

Abstract

During speech recognition of words, a precise and strong detection of start/end points of the words must be ensured, even in very noisy surroundings. Use of a feature with noise-resistant properties is shown wherein for a feature vector, a function of the signal energy is formed as the first feature and a function of the quadratic difference of an LPC (Linear-Predictive-Coding) cepstrum coefficient as a second feature. A check quantity or a maximum function of a distribution function is calculated, which detects the start/end points by comparison with a threshold.

Claims

1. A method of detecting start and end points of words in a signal indicative of speech, with a detected start point indicating a beginning of a word and, at a same time, an end of a nonspeech interval, and a detected end point indicating an end of the word and, at a same time, a beginning of the nonspeech interval, comprising the steps of:

dividing the signal indicative of speech into blocks,

forming a current feature vector from at least two current features, a first of which is a function of a signal energy, and an at least second of which is a function of a quadratic difference between a linear predictive coding (LPC) cepstrum coefficient of a current block and an average LPC cepstrum coefficient,

determining an average feature vector from a predefined number I of blocks containing a nonspeech interval, and updating said average feature vector on an occurrence of each new nonspeech interval, and

using the current feature vector and the average feature vector to determine a check quantity (U) which, compared with a threshold value, provides information as to whether a nonspeech interval or word is present, thus detecting the start and end points.

2. A method as claimed in claim 1 wherein the current feature vector (c(i)) is formed from at least two current features:

##EQU24## a first of which, BMW(i), is a function of the signal energy;

##EQU25## and the at least second of which is a function of the quadratic difference between the current LPC cepstrum coefficient and the average LPC cepstrum coefficient;

##EQU26## and wherein an average feature vector is calculated;

##EQU27## by means of which a covariance matrix is calculated which is used to determine the check quantity (U);

##EQU28## which provides information as to whether a nonspeech interval or word is present.

3. A method as claimed in claim 1 wherein the check quantity (U) is determined by calculating a Mahalanobis distance.

4. A method as claimed in claim 1 wherein the at least second current feature is weighted differently from the first current feature.

5. A method as claimed in claim 1 wherein depending on an ambient noise level, smoothing mechanisms are used.

6. A method as claimed in claim 2 wherein the check quantity (U) is determined by calculating a Mahalanobis distance.

7. A method as claimed in claim 2 wherein the at least second current feature is weighted differently from the first current feature.

8. A method as claimed in claim 3 wherein the at least second current feature is weighted differently from the first current feature.

9. A method as claimed in claim 2 wherein depending on an ambient noise level, smoothing mechanisms and/or adaptive turn-on and turn-off thresholds are used.

10. A method as claimed in claim 3 wherein depending on an ambient noise level, smoothing mechanisms and/or adaptive turn-on and turn-off thresholds are used.

11. A method as claimed in claim 4 wherein depending on an ambient noise level, smoothing mechanisms and/or adaptive turn-on and turn-off thresholds are used.

12. A method of detecting start and end points of words in a signal indicative of speech, comprising the steps of:

dividing the signal indicative of speech into blocks,

forming a current feature vector from at least two current features, a first of which is a function of signal energy, and an at least second of which is a function of a linear predictive coding (LPC) cepstrum coefficient,

determining distribution functions by means of the functions of the current features, and

determining for each block a maximum one of said distribution functions and comparing said maximum one of said distribution functions to a threshold as a measure of whether a nonspeech interval or word occurs between the detected start and end points.

13. A method as claimed in claim 12 wherein the first current feature is a function of the signal energy:

##EQU29## and wherein the at least second feature is a function of the LPC cepstrum coefficients;

##EQU30## where CEPs(k, i) is a short-time average, and wherein distribution functions ##EQU31## and ##EQU32## with ##EQU33## are used to determine a maximum function DMAX(i)=max {Δ

CD(i), Δ

MV(i)} for said step of selecting a maximum one of said distribution functions.

14. A method as claimed in claim 13, wherein logarithmic distribution functions Δ

LCD and Δ

LMV given by Δ

LCD(i)=log₂ (Δ

CD(i)) and Δ

LMV(i)=log₂ (Δ

MV(i)) are used to determine said maximum function in order to determine whether a nonspeech interval or word is present.

15. Program module for detecting the start/end points of words in a signal indicative of speech, comprising:

input/output (I/O) means, responsive to the signal indicative of speech for providing said signal indicative of speech; and

a signal processor, responsive to said signal indicative of speech from said I/O means, for forming a current feature vector for detecting both a start point and an end point, and for forming at least a second feature with noise-resistant properties for said feature vector in which the current feature vector, an average feature vector and a check quantity (U) are formed for detecting both said start point and said end point and for forming a start/end point signal wherein said I/O means is responsive to said start/end point signal for providing said start/end point signal as an output of said program module.