Method of boundary estimation for voice recognition and voice recognition device

• US 5,710,865 A
• Filed: 03/21/1995
• Issued: 01/20/1998
• Est. Priority Date: 03/22/1994
• Status: Expired due to Fees

• Current Assignee: Mitsubishi Denki Kabushiki Kaisha (Mitsubishi Electric Corporation)
• Original Assignee: Mitsubishi Denki Kabushiki Kaisha (Mitsubishi Electric Corporation)
• Inventors: Abe, Yoshiharu
• Primary Examiner(s): MacDonald, Allen R.
• Assistant Examiner(s): Edouard, Patrick N.
• Application Number:
  US08/407,170
• Time in Patent Office: 1,036 Days
• Field of Search: 395/2.57, 395/2.65, 395/2.62, 395/2.64
• US Class Current: 704/248
• CPC Class Codes: G10L 15/04 Segmentation; Word boundary detection
• CPC Class Codes: G10L 15/142 Hidden Markov Models [HMMs]
• CPC Class Codes: G10L 2015/025 Phonemes, fenemes or fenones being the recognition units

Abstract

A method of voice recognition with a boundary binding type in consideration of a kind of a voice boundary estimates a probability where a boundary corresponds to every kind basis of the voice boundary. For a group of parameters within a window of a predetermined time width, the method is to calculate a group of common probability densities independent from the kind of the voice boundary, to multiply the result by first polynomial coefficient and to calculate a first probability where the voice boundary exists in a center of a window, to further multiply the result by a second polynomial coefficient and to calculate a second probability where the voice boundary does not exist in center of a window. A boundary likelihood dependent on a kind of voice boundary is obtained from the first and second probability densities. The voice recognition device compares with a threshold value depending on the kind of phonemic boundary when the time for producing the boundary of the phonemic model series is limited within a region of the phonemic boundary or the neighborhood of the phonemic boundary which is estimated in accordance with a boundary estimating method for the voice corresponding to the kind of phonemic boundary.

Claims

1. A computer-implemented process for estimating a presence of a phonemic boundary within a parametric voice signal, comprising the steps of:

analyzing an input voice to obtain the parametric voice signal;

extracting as a sample each value of the parametric voice signal within a variable position window of a predetermined time width;

calculating for the sample extracted at each position of the window a common probability density which represents a probability that the phonemic boundary lies at a center of the window;

calculating a first probability density estimating whether a particular boundary type lies at the center of the window, by multiplying the common probability density calculated by a first polynomial coefficient;

calculating a second probability density estimating whether the particular boundary type does not lie at the center of the window, by multiplying the common probability density calculated by a second polynomial coefficient; and

producing a signal indicative of the estimate of the presence of the phonemic boundary as a function of the first and second probability densities.

3. A computer implemented process as recited in claim 1, the step of analyzing further comprising:

detecting when an input voice signal lies within a voice zone based upon a power calculation.

4. A computer implemented process as recited in claim 3, further comprising the steps of:

performing a hidden Markov model calculation based on the estimate of the presence of a phonemic boundary, the parametric voice signal sample and a hidden Markov model parameter.

5. A computer implemented process as recited in claim 4, further comprising the step of:

computing an optimum state series based upon the hidden Markov model computation.

6. A computer implemented process as recited in claim 5, wherein the step of computing an optimum state series further comprises the steps of:

establishing a boundary likelihood classification relative to a state transition in the hidden Markov model;

comparing the boundary likelihood classification to a variable threshold related to the state transition;

changing state in the hidden Markov model in accordance with the state transition when the boundary likelihood classification exceeds the variable threshold; and

varying the variable threshold to have high values corresponding to easy to detect phonemic boundaries and low values corresponding to difficult to detect phonemic boundaries.

2. A voice recognition device including

a feature extracting means for analyzing an input voice and converting into a feature parameter of time series;

a boundary detecting means for detecting a region of a phonemic boundary or a neighborhood of the phonemic boundary in the input voice from the feature parameter of time series, the boundary detecting means first comparing a probability that the phonemic boundary or the neighborhood of the phonemic boundary exists within a window having a predetermined time width to a probability that the phonemic boundary or the neighborhood of the phonemic boundary does not exist within the window and comparing a result of the first comparison to a threshold having a value that varies dependent on a particular type of phonemic boundary that is detected;

a model calculating means in which a time of producing each boundary of a plurality of phonemic model series is limited within a region of the phonemic boundary or the neighborhood of the phonemic boundary which the boundary detecting means detects when preparing the model series, the phonemic model series being prepared corresponding to the feature parameter; and

a phonemic series converting means for selecting a suitable phonemic model series corresponding to the input voice from a result of the model calculating means.

22. The voice recognition device of claim 2, wherein the value of the threshold increases when the particular type of phonemic boundary is easy to detect and wherein the value of the threshold decreases when the particular type of phonemic boundary is difficult to detect.

23. The voice recognition device of claim 2, wherein the boundary detecting means includes:

common probability calculating means for calculating a common probability density that the phonemic boundary or the neighborhood of the phonemic boundary exists within the window independently of the particular type of phonemic boundary;

first means for multiplying the common probability density by a first polynomial coefficient to estimate the probability that the phonemic boundary or the neighborhood of the phonemic boundary exists within the window; and

second means for multiplying the common probability density by a second polynomial coefficient to estimate the probability that the phonemic boundary or the neighborhood of the phonemic boundary does not exist within the window.

7. A method of producing a signal representing an estimate of the presence at a time location in a voice signal of a phonemic boundary comprising the steps of:

receiving the voice signal into an analysis system;

analyzing the voice signal to produce a parametric voice signal representing the voice signal as a time series of parameter groups;

extracting as a sample each parameter group within a variable position window of a predetermined time width;

calculating for the sample extracted at each position of the window a common probability density which represents a probability that the phonemic boundary lies at a center of the window;

calculating a first probability density estimating whether a particular boundary type lies at the center of the window, by multiplying the common probability density calculated by a first polynomial coefficient;

calculating a second probability density estimating whether the particular boundary type does not lie at the center of the window, by multiplying the common probability density calculated by a second polynomial coefficient; and

producing a signal indicative of the estimate of the presence of the phonemic boundary as a function of the first and second probability densities.

8. A method of producing as recited in claim 7, the step of analyzing further comprising:

detecting when an input voice signal lies within a voice zone based upon a power calculation.

9. A method of producing as recited in claim 8, farther comprising the steps of:

performing a hidden Markov model calculation based on the estimate of the presence of a phonemic boundary, the parametric voice signal sample and a hidden Markov model parameter.

10. A method of producing as recited in claim 9, further comprising the step of:

computing an optimum state series based upon the hidden Markov model computation.

11. A method of producing as recited in claim 10, wherein the step of computing an optimum state series further comprises the steps of:

establishing a boundary likelihood classification relative to a state transition in the hidden Markov model;

comparing the boundary likelihood classification to a variable threshold related to the state transition;

changing state in the hidden Markov model in accordance with the state transition when the boundary likelihood classification exceeds the variable threshold; and

varying the variable threshold to have high values corresponding to easy to detect phonemic boundaries and low values corresponding to difficult to detect phonemic boundaries.

12. A method of producing a signal representing an estimate of the presence at a time location in a voice signal of a phonemic boundary comprising the steps of:

receiving the voice signal into an analysis system;

analyzing the voice signal to produce a parametric voice signal representing the voice signal as a time series of parameter groups;

extracting as a sample each parameter group within a window of a predetermined time width, the window centered at a time location;

calculating from the time series of parameter groups a probability that a phonemic boundary lies at the time location;

multiplying the probability calculated by a first polynomial coefficient to form a probability that a particular type of phonemic boundary lies at the time location;

multiplying the probability calculated by a second polynomial coefficient to form a probability that the particular type of phonemic boundary does not lie at the time location; and

producing a signal indicative of the estimate of the presence of the phonemic boundary as a function of the first and second probability densities.

13. A method of producing a signal as recited in claim 12, the step of analyzing further comprising:

detecting when an input voice signal lies within a voice zone based upon a power calculation.

14. A method of producing a signal as recited in claim 13, further comprising the steps of:

performing a hidden Markov model calculation based on the estimate of the presence of a phonemic boundary, the parametric voice signal sample and a hidden Markov model parameter.

15. A method of producing a signal as recited in claim 14, further comprising the step of:

computing an optimum state series based upon the hidden Markov model computation.

16. A method of producing a signal as recited in claim 15, wherein the step of computing an optimum state series further comprises the steps of:

establishing a boundary likelihood classification relative to a state transition in the hidden Markov model;

comparing the boundary likelihood classification to a variable threshold related to the state transition;

changing state in the hidden Markov model in accordance with the state transition when the boundary likelihood classification exceeds the variable threshold; and

varying the variable threshold to have high values corresponding to easy to detect phonemic boundaries and low values corresponding to difficult to detect phonemic boundaries.

17. Apparatus for estimating a likelihood of presence of a phonemic boundary at a time location within a voice signal, comprising:

a voice signal analyzing device having an input receiving the voice signal, having an output carrying a parametric voice signal, the voice analyzing device producing the parametric voice signal as a time series of parameter groups;

a segment extraction device connected to receive from the voice signal analyzing device the parametric voice signal, and producing at an output a signal representing a segment of the voice signal including a time series of parameter groups falling within a variable position window of a predetermined time width, the window having a center at a time location;

a first probability calculating device, receiving the segment of the voice signal from the segment extraction device and producing an output signal representing a probability that the phonemic boundary lies at the time location;

a second probability calculating device, receiving the probability from the first probability calculating device, and multiplying the probability received by a first polynomial coefficient to produce a signal indicative of likelihood of presence of a particular type of phonemic boundary at the time location;

a third probability calculating device, receiving the probability from the first probability calculating device, and received by a second polynomial coefficient to produce a signal indicative of likelihood that a particular type of phonemic boundary does not lie at the time location.

18. Apparatus as recited in claim 17, the voice signal analyzing device further comprising:

a comparator receiving the input voice signal and determining when the input voice signal lies within a voice zone based upon a power calculation.

19. Apparatus as recited in claim 18, further comprising:

a device having inputs receiving the estimate of the presence of a phonemic boundary, the parametric voice signal sample and a hidden Markov model parameter and producing an output representing a hidden Markov model.

20. Apparatus as recited in claim 19, further comprising:

a state machine receiving as an input the hidden Markov model computation and producing as an output an optimum state series.

21. Apparatus as recited in claim 20, wherein the state machine further comprises:

means for establishing a boundary likelihood classification relative to a state transition in the hidden Markov model;

means for comparing the boundary likelihood classification to a variable threshold related to the state transition;

means for changing state in the hidden Markov model in accordance with the state transition when the boundary likelihood classification exceeds the variable threshold; and

means for varying the variable threshold to have high values corresponding to easy to detect phonemic boundaries and low values corresponding to difficult to detect phonemic boundaries.