Granted Patent B2

US 8,566,084 · App. 13/150,804 · Granted Oct 22, 2013

Inventors: Osamu Ichikawa (Kanagawa-ken, JP); Masafumi Nishimura (Kanagawa-ken, JP)

Assignee: Nuance Communications, Inc.

Abstract

Claims (45)

1. A speech signal processing system, comprising:

a framing unit configured to divide an input speech signal into frames, wherein a pair of consecutive frames has a frame shift length equal to at least one fundamental period of the input speech signal, and the pair of consecutive frames has an overlap equal to at least a predetermined length;

a discrete Fourier transform unit configured to apply a discrete Fourier transform to each of the frames and output a spectrum of the input speech signal;

a Cross-Power Spectrum Phase (CSP) coefficient calculation unit configured to calculate a CSP coefficient of the spectrum of each pair of consecutive frames;

a phase information obtaining unit configured to receive an output from the CSP coefficient calculation unit, and to search a predetermined search range for a maximum CSP coefficient value for each pair of consecutive frames, so as to obtain an index that maximizes the CSP coefficient value for each pair of consecutive frames;

a time-series data generation unit configured to arrange the maximum CSP coefficient values for pairs of consecutive frames in the order in which the frames appear so as to generate time-series data; and

a vowel identification unit configured to use the time series data of the maximum CSP values to identify vowel parts of the input speech signal.

2. The system of claim 1 , wherein the pair of consecutive frames comprises an earlier frame and a later frame, wherein the predetermined search range is a range from −L/2 to +L/2 with respect to a start position of the later frame, and wherein L is 10 milliseconds.

3. The system of claim 1 , wherein the frame shift length is 10 milliseconds.

4. The system of claim 1 , further comprising:

a vowel enhancement unit configured to match, for each pair of consecutive frames, phases of spectra of the pair of frames on the basis of the value of an index that maximizes the CSP coefficient obtained for the pair of frames in the predetermined search range, and to add the spectra together; and

a feature vector calculation unit configured to receive an output from the vowel enhancement unit as an input, extract a speech feature from the input, and output the speech feature along with the time-series data of the maximum values of the CSP coefficient as a speech feature vector.

5. The system of claim 1 , further comprising:

a masking unit configured to calculate, for each pair of consecutive frames, a masker that weakens a frequency bin in which the signs of spectra of the pair do not match when the spectra are compared with each other in a position in which a phase lag between the spectra is minimized on the basis of the value of an index that maximizes the CSP coefficient obtained for the pair in the predetermined speech range, and to apply the masker to the spectrum of one frame of the pair; and

a feature vector calculation unit configured to receive an output from the masking unit as an input, extract a speech feature from the input, and output the speech feature along with the time-series data of the maximum values of the CSP coefficient as a speech feature vector.

6. A method of processing a speech signal, the method comprising:

dividing, using a framing unit, an input speech signal into frames, wherein a pair of consecutive frames has a frame shift length equal to at least one fundamental period of the input speech signal, and the pair of consecutive frames has an overlap equal to at least a predetermined length;

applying, using a discrete Fourier transform unit, a discrete Fourier transform to each of the frames and outputting a spectrum of the input speech signal;

calculating, using a Cross-Power Spectrum Phase (CSP) coefficient calculation unit, a CSP coefficient of the spectrum of each pair of consecutive frames;

receiving, using a phase information obtaining unit, an output from the CSP coefficient calculation unit, and searching a predetermined search range for a maximum CSP coefficient value for each pair of consecutive frames, so as to obtain an index that maximizes the CSP coefficient value for each pair of consecutive frames;

arranging, using a time-series data generation unit, the maximum CSP coefficient values for pairs of consecutive frames in the order in which the frames appear so as to generate time-series data; and

using the time series data of the maximum CSP values to identify vowel parts of the input speech signal.

7. The method of claim 6 , wherein the pair of consecutive frames comprises an earlier frame and a later frame, wherein the predetermined search range is a range from −L/2 to +L/2 with respect to a start position of the later frame, and wherein L is 10 milliseconds.

8. The method of claim 6 , wherein the frame shift length is 10 milliseconds.

9. The method of claim 6 , further comprising:

matching, using a vowel enhancement unit, for each pair of consecutive frames, phases of spectra of the pair of frames on the basis of the value of an index that maximizes the CSP coefficient obtained for the pair of frames in the predetermined search range, and adding the spectra together; and

receiving, using a feature vector calculation unit, an output from the vowel enhancement unit as an input, extracting a speech feature from the input, and outputting the speech feature along with the time-series data of the maximum values of the CSP coefficient as a speech feature vector.

10. The method of claim 6 , further comprising:

calculating, using a masking unit, for each pair of consecutive frames, a masker that weakens a frequency bin in which the signs of spectra of the pair do not match when the spectra are compared with each other in a position in which a phase lag between the spectra is minimized on the basis of the value of an index that maximizes the CSP coefficient obtained for the pair in the predetermined search range, and applying the masker to the spectrum of one frame of the pair; and

receiving, using a feature vector calculation unit, an output from the masking unit as an input, extracting a speech feature from the input, and outputting the speech feature along with the time-series data of the maximum values of the CSP coefficient as a speech feature vector.

11. A computer program product comprising a non-transitory, computer readable storage medium having computer readable instructions stored thereon that, when executed by a computer, implement a method of processing a speech signal, the method comprising:

dividing, using a framing unit, an input speech signal into frames, wherein a pair of consecutive frames has a frame shift length equal to at least one fundamental period of the input speech signal, and the pair of consecutive frames has an overlap equal to at least a predetermined length;

applying, using a discrete Fourier transform unit, a discrete Fourier transform to each of the frames and outputting a spectrum of the input speech signal;

calculating, using a Cross-Power Spectrum Phase (CSP) coefficient calculation unit, a CSP coefficient of the spectrum of each pair of consecutive frames;

receiving, using a phase information obtaining unit, an output from the CSP coefficient calculation unit, and searching a predetermined search range for a maximum CSP coefficient value for each pair of consecutive frames, so as to obtain an index that maximizes the CSP coefficient value for each pair of consecutive frames;

arranging, using a time-series data generation unit, the maximum CSP coefficient values for pairs of consecutive frames in the order in which the frames appear so as to generate time-series data; and

using the time series data of the maximum CSP values identify vowel parts of the input speech signal.

12. The computer program product of claim 11 , wherein the pair of consecutive frames comprises an earlier frame and a later frame, wherein the predetermined search range is a range from −L/2 to +L/2 with respect to a start position of the later frame, and wherein L is 10 milliseconds.

13. The computer program product of claim 11 , wherein the frame shift length is 10 milliseconds.

14. The computer program product of claim 11 , wherein the method further comprises:

matching, using a vowel enhancement unit, for each pair of consecutive frames, phases of spectra of the pair of frames on the basis of the value of an index that maximizes the CSP coefficient obtained for the pair of frames in the predetermined search range, and adding the spectra together; and

receiving, using a feature vector calculation unit, an output from the vowel enhancement unit as an input, extracting a speech feature from the input, and outputting the speech feature along with the time-series data of the maximum values of the CSP coefficient as a speech feature vector.

15. The computer program product of claim 11 , wherein the method further comprises:

calculating, using a masking unit, for each pair of consecutive frames, a masker that weakens a frequency bin in which the signs of spectra of the pair do not match when the spectra are compared with each other in a position in which a phase lag between the spectra is minimized on the basis of the value of an index that maximizes the CSP coefficient obtained for the pair in the predetermined search range, and applying the masker to the spectrum of one frame of the pair; and

receiving, using a feature vector calculation unit, an output from the masking unit as an input, extracting a speech feature from the input, and outputting the speech feature along with the time-series data of the maximum values of the CSP coefficient as a speech feature vector.

Assignments (9)

Priority Claims (1)

Continuity (1)