Restoration of high-order Mel Frequency Cepstral Coefficients

1. A method for estimating high-order Mel Frequency Cepstral Coefficients, the method comprising:

a) in an MFCC vector of length N having L low-order coefficients (LOC), initializing any of N−L high-order coefficients (HOC) of said MFCC vector to a predetermined value, thereby forming a candidate MFCC vector;

b) synthesizing a speech signal frame from said candidate MFCC vector and a pitch value using a harmonic model of voiced speech for parametric representation of said speech frame;

c) computing an N-dimensional MFCC vector from said synthesized frame, thereby producing an output MFCC vector;

d) performing said steps b)-c) up to a predetermined number of additional iterations, wherein the HOC of said output MFCC vector of a given iteration is appended to said LOC to form a new candidate MFCC vector for the next iteration; and

e) estimating for each of a plurality of iterations of steps b)-c) any of the parameters of said model from said candidate MFCC vector and said pitch value for said iteration.

2. A method according to claim 1 wherein said computing step comprises calculating using a harmonic model of voiced speech for parametric representation of said speech frame, wherein at each of a plurality of iterations of steps b)-c) said output MFCC vector is computed from said harmonic model parameters estimated at that iteration.

3. A method for estimating high-order Mel Frequency Cepstral Coefficients, the method comprising:

a) convening a truncated L-dimensional MFCC vector of low-order coefficients (LOC) to an N-dimensional binned spectrum;

b) initializing N−L high-order coefficients (HOC) using predetermined values;

c) computing an N-dimensional binned spectrum corresponding to said HOC;

d) calculating a composite binned spectrum from both of said binned spectra using coordinate-wise multiplication;

e) estimating at least one harmonic model parameter from said composite binned spectrum and a pitch frequency, thereby producing a basis bins matrix and basis function mixing coefficients;

f) synthesizing a new binned spectrum by multiplying said basis bins matrix by the vector of said basis function mixing coefficients;

g) regularizing said synthesized bins; and

h) convening said regularized synthesized bins to HOC, thereby estimating said HOC.

4. A method according to claim 3 wherein said convening step a) comprises converting using an N-dimensional Inverse Discrete Cosine Transform (IDCT) followed by an antilog operation.

5. A method according to claim 4 wherein said converting step a) comprises appending to said truncated MFCC vector an N−L-dimensional vector of zero-valued coordinates.

6. A method according to claim 3 wherein said initializing step b) comprises initializing using zero values, and wherein each coordinate of said binned spectrum vector corresponding to said HOC is set equal to 1.

7. A method according to claim 3 wherein said initializing step b) comprises:

preparing a set of HOC vectors, wherein each vector corresponds to a predetermined range of pitch values;

determining the range into which a provided pitch value fits;

selecting from among said HOC vectors a vector that corresponds to said range; and

initializing said HOC with said selected vector.

8. A method according to claim 3 wherein said computing step c) comprises logically preceding said initialized HOC vector by N−L zeros.

9. A method according to claim 3 wherein said estimating step e) comprises modeling at least one harmonic amplitude A k as a linear combination of N basis functions {BF i } sampled at a plurality of pitch frequency multiples

as

⁢

⁢

A

k

=

∑

i

=

1

N

⁢

b

i

·

B

⁢

⁢

F

i

⁡

(

k

⁢

⁢

F

p

)

.

10. A method according to claim 3 and further comprising performing a transformation √{square root over (B i ·S i )} of each coordinate B i of said composite binned spectrum, wherein S i is a sum of the values of the i th Mel-filter, where said input LOC was produced using a Power Short Time Spectrum.

11. A method according to claim 3 wherein said regularizing step g) comprises:

identifying any coordinates of said synthetic vector whose value does not exceed a predefined threshold; and

setting any of said identified coordinates equal to a regularization value R=a·B av , wherein B av is an average of the coordinate values of said synthetic vector and a is predefined value.

12. A method according to claim 3 wherein said converting step h) comprises:

applying a logarithm to synthetic vector; and

performing a Discrete Cosine Transform on said synthetic vector.

13. A method according to claim 3 and further comprising:

i) computing a new binned spectrum corresponding to said HOC vector; and

j) performing steps d)-h) using said new binned spectrum corresponding to said HOC vector.

14. A method according to claim 13 and further comprising:

performing steps i)-j) a plurality of times until a predefined number of iterations is reached; and

if said predefined number of iterations has been reached, concatenating said estimated HOC with said LOC.

15. A method according to claim 3 wherein said estimating step e) comprises:

sampling a plurality of basis functions at a plurality of pitch frequency multiples;

calculating a spectral envelope for each sampled basis function by convolution with a Fourier transform of a windowing function;

applying Mel filters to each of said spectral envelopes, resulting in an N-dimensional basis bins vector BB 1 ;

composing an N by N basis bins matrix BB having said basis bins vectors BB i as its columns;

computing an equation matrix Q as Q=BB T *BB+ε*I where I is a unit matrix and ε is a regularization factor;

applying LU-decomposition to said equation matrix Q;

computing an equation right side vector V as V BB T *B, where B is a column vector of bin value inputs to harmonic model parameter estimation; and solving a matrix equation Q*b=V in b using the LU representation of said equation matrix Q.

16. A method according to claim 15 wherein said computing an equation matrix step comprises computing said regularization factor as 0.001 multiplied by the average of the BB T *BB matrix elements residing at the main diagonal.

17. A Distributed Speech Recognition system employing MFCC vector HOC estimation, the system comprising:

speech recognition front-end apparatus operative to extract from each frame of an input speech signal a LOC, a pitch value, and a voicing class;

HOC restoration apparatus operative to:

form a candidate MFCC vector from said LOC and a plurality of HOC,

synthesize a speech signal frame from said candidate MFCC vector and said pitch value,

apply speech recognition front-end processing to said synthesized frame, thereby producing an output MFCC vector, and

sample basis functions at a plurality of pitch frequencies and compute a plurality of mixing coefficients;

speech recognition back-end apparatus operative to produce text from a plurality of said output MFCC vectors;

speech reconstruction apparatus operative to synthesize speech from plurality of said output MFCC vectors, said pitch values, and said voicing class values; and

harmonic amplitudes modeling apparatus operative to calculate harmonic amplitudes from said basis functions and mixing coefficients,

wherein said speech reconstruction apparatus is operative to synthesize said speech from a plurality of said output MFCC vectors, said pitch values, said voicing class values, and said harmonic amplitudes.

18. A computer program embodied on a computer-readable medium, the computer program comprising:

a) a first code segment operative for an MFCC vector of length N having L low-order coefficients (LOC) to initialize any of N−L high-order coefficients (HOC) of said MFCC vector to a predetermined value, thereby forming a candidate MFCC vector;

b) a second code segment operative to synthesize a speech signal frame from said candidate MFCC vector and a pitch value using a harmonic model of voiced speech for parametric representation of said speech frame;

c) a third code segment operative to compute an N-dimensional MFCC vector from said synthesized frame, thereby producing an output MFCC vector;

d) a fourth code segment operative to execute said second and third code segments up to a predetermined number of iterations, wherein the HOC of said output MFCC vector of a given iteration is appended to said LOC to form a new candidate MFCC vector for the next iteration; and

e) a fifth code segment operative to estimate for each of a plurality of said iterations any of the parameters of said model from said candidate MFCC vector and said pitch value for said iteration.

19. A computer program embodied on a computer-readable medium, the computer program comprising:

a first code segment operative to convert a truncated L-dimensional MFCC vector of low-order coefficients (LOC) to an N-dimensional binned spectrum;

a second code segment operative to initialize N−L high-order coefficients (HOC) using predetermined values;

a third code segment operative to compute an N-dimensional binned spectrum corresponding to said HOC;

a fourth code segment operative to calculate a composite binned spectrum from both of said binned spectra using coordinate-wise multiplication;

a fifth code segment operative to estimate at least one harmonic model parameter from said composite binned spectrum and a pitch frequency, thereby producing a basis bins matrix and basis function mixing coefficients;

a sixth code segment operative to synthesize a new binned spectrum by multiplying said basis bins matrix by the vector of said basis function mixing coefficients;

a seventh code segment operative to regularize said synthesized bins; and

a eighth code segment operative to convert said regularized synthesized bins to HOC, thereby estimating said HOC.