# US 2004/0083096 A1 · Method and apparatus for gradient-descent based window optimization for linear prediction analysis

> Agents: this data is available as JSON from the patentgrep API and MCP server — see https://patentgrep.com/llms.txt

- Number: US 2004/0083096 A1
- Document ID: US20040083096A1
- Status: Published application
- Kind: A1
- Application number: 10/282,966
- Filed: 2002-10-29
- Published: 2004-04-29
- Earliest claimed: 2002-10-29
- Assignees: Ntt Docomo, Inc.
- Inventors: Wai C. Chu
- Web page: https://patentgrep.com/patent/US20040083096A1

## CPC

- G10L 19/022: Blocking, i.e. grouping of samples in time; Choice of analysis windows; Overlap factoring
- G10L 19/04: using predictive techniques
- G10L 25/12: the extracted parameters being prediction coefficients
- G10L: SPEECH ANALYSIS TECHNIQUES OR SPEECH SYNTHESIS; SPEECH RECOGNITION; SPEECH OR VOICE PROCESSING TECHNIQUES; SPEECH OR AUDIO CODING OR DECODING

## Abstract

The shape of windows used during linear predictive analysis can be optimized through the use of gradient-descent based window optimization procedures. Window optimization may be achieved fairly precisely through the use of a primary optimization procedure, or less precisely through the use of an alternate optimization procedure. Both optimization procedures use the principle of gradient-descent to find a window sequence that will either minimize the prediction error energy or maximize the segmental prediction gain. However, the primary optimization procedure uses a Levinson-Durbin based algorithm to determine the gradient while the alternate optimization procedure uses an estimate of the gradient based on the basic definition of a derivative. These optimization procedures can be implemented as computer readable software code. Additionally, the optimization procedures may be implemented in a window optimization device which generally includes a window optimization unit and may also include an interface unit.

## Claims

### Claim 1

An optimization procedure for optimizing window sequences used in linear prediction analysis, comprising:
an initialization procedure, wherein the initialization procedure assumes an initial window sequence, and defines the initial window sequence as a window sequence;
a gradient-descent procedure, wherein the gradient descent procedure:
determines an updated window sequence, and defines the updated window sequence as the window sequence
determines a gradient of a prediction error energy wherein the gradient is determined using the window sequence; and
a stop procedure, wherein the stop procedure determines if a threshold is met, wherein if the threshold is not met, the gradient-descent procedure and the stop procedure are repeated until the threshold is met.

### Claim 2 (depends on 1)

An optimization procedure, as claimed in claim 1, wherein the initialization procedure computes an initial prediction error energy and a derivative of the initial prediction error energy using the initial widow sequence and a Levinson-Durbin initialization procedure.

### Claim 3 (depends on 1)

An optimization procedure, as claimed in claim 1, wherein the gradient descent procedure determines the gradient of the prediction error energy using the recursion routine of a Levinson-Durbin algorithm.

### Claim 4 (depends on 1)

An optimization procedure, as claimed in claim 1, wherein the initialization procedure computes an initial prediction error energy using linear prediction analysis.

### Claim 5 (depends on 1)

An optimization procedure, as claimed in claim 1, wherein the gradient descent procedure estimates the gradient of the prediction error energy using an estimated based on a definition of a partial derivative.

### Claim 6

A method for optimizing a window in linear prediction analysis of a speech signal, comprising:
assuming a initial window sequence, wherein the initial window sequence is a window sequence, wherein the window sequence comprises a plurality of window samples and wherein the length of the window sequence is N;
determining a gradient of a prediction error energy of the speech signal, wherein the speech signal is windowed by the initial window sequence;
updating the window sequence to create a next window sequence, wherein the next window sequence becomes the window sequence;
determining a gradient of a new prediction error energy of the speech signal, wherein the speech signal is windowed by the window sequence; and
determining whether a threshold has been reached; wherein if the threshold has not been reached, repeating the steps of updating the window to create the next window sequence, determining the gradient of the prediction error energy of the speech signal windowed by the window sequence wherein the next window sequence becomes the window sequence, and determining whether the threshold has been reached, until the threshold is reached.

### Claim 7 (depends on 6)

A window optimization method, as claimed in claim 6, wherein assuming the initial window sequence comprises assuming a rectangular window sequence.

### Claim 8 (depends on 6)

A window optimization method, as claimed in claim 6, wherein determining the gradient of the prediction error energy of the speech signal comprises using a Levinson-Durbin initialization routine.

### Claim 9 (depends on 8)

A window optimization method, as claimed in claim 8, wherein determining the gradient of the prediction error energy of the speech signal using a Levinson-Durbin initialization routine comprises:
defining a time lag l, wherein l equals zero;
determining an initial autocorrelation value with respect to each window sample of the initial window R[l], for l=0;
determining a partial derivative of the initial autocorrelation value with respect to each window sample of the initial window sequence, wherein a partial derivative of the initial autocorrelation value with respect to each window sample of the initial window sequence is indicated by
25 ∂ R ⁡ [ l ] ∂ w ⁡ [ n ]
wherein l=0; and
determining a prediction error energy and a partial derivative of the prediction error energy as a function of the initial autocorrelation value with respect to each window sample of the initial window, wherein each of the prediction error energies are indicated by Jo and each of the partial derivatives of the prediction error energy is indicated by
26 ∂ J l ∂ w ⁡ [ n ]
wherein l=0.

### Claim 10 (depends on 9)

A window optimization method, as claimed in claim 9, wherein determining R[l] for l=0 comprises determining R[l] for l=0 as a function of the window sequence and the input signal and according to an equation
27 R ⁡ [ l ] = ∑ k = 1 N - 1 ⁢ ⁢ w ⁡ [ k ] ⁢ s ⁡ [ k ] ⁢ w ⁡ [ k - l ] ⁢ s ⁡ [ k - l ] ⁢ ⁢ for ⁢ ⁢ l = 0.

### Claim 11 (depends on 9)

A window optimization method, as claimed in claim 9, wherein determining
28 ∂ R ⁡ [ l ] ∂ w ⁡ [ n ]
for l=0 comprises determining
29 ∂ R ⁡ [ l ] ∂ w ⁡ [ n ]
for l=0 according to known values.

### Claim 12 (depends on 6)

A window optimization method, as claimed in claim 6, wherein updating the window sequence comprises defining the next window sequence as a function of a step size parameter.

### Claim 13 (depends on 6)

A window optimization method, as claimed in claim 6, wherein determining the gradient of the new prediction error energy of the speech signal comprises using a Levinson-Durbin recursion routine.

### Claim 14 (depends on 13)

A window optimization method, as claimed in claim 13, wherein determining the gradient of the new prediction error energy of the speech signal using the Levinson-Durbin recursion routine, comprises:
determining a linear predictive coefficient and a partial derivatives\ of the linear predictive coefficients for each of the window samples of the window sequence, wherein each of the linear predictive coefficients are indicated by an index i as ai and each of the partial derivatives of the linear predictive coefficients are indicated by
30 ∂ a i ∂ w ⁡ [ n ] ;
determining a prediction error sequence as a function of the speech signal windowed by the window sequence and the linear predictive coefficients, wherein the prediction error sequence comprises a new prediction energy estimate for each of the window samples of the window sequence;
determining a partial derivative of the new prediction energy estimate with respect to each of the window samples of the window sequence, wherein the partial derivative of the new prediction energy estimate with respect to each of the window samples of the window sequence is indicated by
31 ∂ J ∂ w ⁡ [ n ] .

### Claim 15 (depends on 9)

A window optimization method, as claimed in claim 9, wherein determining the linear predictive coefficients and the partial derivatives of the linear predictive coefficients for each of the plurality of window samples of the window sequence comprises using a Levinson-Durbin algorithm.

### Claim 16 (depends on 15)

A window optimization method, as claimed in claim 15, wherein using the Levinson-Durbin algorithm comprises:
incrementing the time lag l, by defining l according to an equation l=l+1;
determining an l-order autocorrelation value with respect to each of the plurality of window samples of the window, wherein each of the l-order autocorrelation values is indicated by Rμl;
determining a partial derivative of each of the l-order autocorrelation values with respect to each of the window samples of the window sequence, wherein each of the l-order autocorrelation values is indicated by
32 ∂ R ⁡ [ l ] ∂ w ⁡ [ n ] ;
calculating the linear predictive coefficients and the partial derivative of each of the linear predictive coefficients with respect to each of the window samples of the window sequence, wherein each of the linear predictive coefficients are indicated by an index i as ai and each of the partial derivatives of the linear predictive coefficients are indicated by
33 ∂ a i ∂ w ⁡ [ n ] ;
; and
determining if l equals an order M, wherein if l does not equal the order M, repeating the steps of incrementing the time lag l by defining l according to and equation l=l+1; determining R[l]; determining
34 ∂ R ⁡ [ l ] ∂ w ⁡ [ n ]
calculating the linear predictive coefficients and the partial derivatives of the linear predictive coefficients with respect to each of the window samples of the window sequence; and determining if l equals an order M until l equals an order M.

### Claim 17 (depends on 16)

A window optimization method, as claimed in claim 16, wherein determining R[l] comprises determining R[l] as a function of a plurality of indices k, the window length N, the plurality of speech signal samples s[k], and the plurality of window samples w[k] of the window sequence, wherein R[l] is defined by an equation
35 R ⁡ [ l ] = ∑ k = l N - 1 ⁢ ⁢ w ⁡ [ k ] ⁢ s ⁡ [ k ] ⁢ w ⁡ [ k - l ] ⁢ s ⁡ [ k - l ] .

### Claim 18 (depends on 16)

A window optimization method, as claimed in claim 16, wherein determining
36 ∂ R ⁡ [ l ] ∂ w ⁡ [ n ]
comprises determining
37 ∂ R ⁡ [ l ] ∂ w ⁡ [ n ]
according to known values.

### Claim 19 (depends on 16)

A window optimization method, as claimed in claim 16, wherein calculating ai and
38 ∂ a i ∂ w ⁡ [ n ]
comprises:
determining a reflection coefficient for each of the window samples of the window sequences and a partial derivative of each of the reflection coefficients for each of the window samples of the window sequences, wherein each of the reflection coefficients are indicated by kl and the partial derivative of each of the reflection coefficients is indicated by
39 ∂ k l ∂ w ⁡ [ n ] ;
determining at least two update functions for each window sample of the window sequence and a partial derivative of each of the at least two update functions for each window sample of the window sequence, wherein the at least two update functions are indicated by a(l)i=−kl and a(l)i=ai(l−1)−klal-i(l−1) and the partial derivative of each of the at least two update functions is indicated by
40 ∂ a i ( l ) ∂ w ⁡ [ n ] = - ∂ k l ∂ w ⁡ [ n ]
and
41 ∂ a i ( l ) ∂ w ⁡ [ n ] = ∂ a i ( l - 1 ) ∂ w ⁡ [ n ] - a l - i ( l - 1 ) ⁢ ∂ k l ∂ w ⁡ [ n ] - k I ⁢ ∂ a l - i ( l - 1 ) ∂ w ⁡ [ n ] ;
determining an l-order partial derivative of the linear predictive coefficients with respect to each window sample of the window sequence; and
determining if l equals M, wherein if l does not equal M, updating the l-order prediction error energy and the partial derivative of the prediction error energy, wherein the prediction error energy is indicated by Jl and the partial derivative of the prediction error energy is indicated by
42 ∂ J l ∂ w ⁡ [ n ] ,
and repeating determining the at least two update functions and the partial derivative of each of the at least two update functions, for each window sample of the window sequence and determining if l equals M until l equals M; wherein when l equals M, defining the linear predictive coefficients according to an equation al=al(M) and defining the partial derivative of the linear predictive coefficients according to an equation
43 ∂ a i ∂ w ⁡ [ n ] = ∂ a i ( M ) ∂ w ⁡ [ n ]
for each window sample of the window sequence.

### Claim 20 (depends on 16)

A window optimization method, as claimed in claim 16, wherein determining the partial derivative of each of the reflection coefficients kl with respect to each of the window samples of the window sequence comprises defining the partial derivative of each of the reflection coefficients kl with an equation
44 ∂ k l ∂ w ⁡ [ n ] = 1 J l - 1 ⁢ ( ∂ R ⁡ [ l ] ∂ w ⁡ [ n ] - R ⁡ [ l ] J l - 1 ⁢ ∂ J l - 1 ∂ w ⁡ [ n ] + ∑ i = 1 l - 1 ⁢ a i ( l - 1 ) ⁢ ∂ R ⁡ [ l - i ] ∂ w ⁡ [ n ] + R ⁡ [ l - i ] ⁢ ∂ a i ( l - 1 ) ∂ w ⁡ [ n ] - a i ( l - 1 ) ⁢ R ⁡ [ l - i ] J l - 1 ⁢ ∂ J l - 1 ∂ w ⁡ [ n ] ) ;

### Claim 21 (depends on 16)

A window optimization method, as claimed in claim 16, wherein defining the l-order partial derivative of the linear prediction coefficients comprises defining the l-order partial derivative of the linear prediction coefficients according to an equation,
45 ∂ a i ( l ) ∂ w ⁡ [ n ] = ∂ a i ( l - 1 ) ∂ w ⁡ [ n ] - a l - i ( l - 1 ) ⁢ ∂ k l ∂ w ⁡ [ n ] = k l ⁢ ∂ a l - i ( l - 1 ) ∂ w ⁡ [ n ] ,
for i=1, 2, l−1-1.

### Claim 22 (depends on 19)

A window optimization method, as claimed in claim 19, wherein updating the l-order prediction error energy and the partial derivative of the prediction error energy further comprises:
updating Jl, wherein Jl is updated according to an equation
Jl=Jl−1(1−k2);
and
updating
46 ∂ J l ∂ w ⁡ [ n ] ,
wherein
47 ∂ J l ∂ w ⁡ [ n ]
is updated according to an equation
48 ∂ J l ∂ w ⁡ [ n ] = ( 1 - k l 2 ) ⁢ ∂ J l - 1 ∂ w ⁡ [ n ] - 2 ⁢ k l ⁢ J l - 1 ⁢ ∂ k l ∂ w ⁡ [ n ] .

### Claim 23 (depends on 14)

A window optimization method, as claimed in claim 14, wherein, determining the prediction error sequence as a function of the speech signal windowed by the window sequence and the linear predictive coefficients, comprises: determining the prediction error sequence e[n] over a synthesis interval n wherein n ε[n1, n2], as defined by an equation,
49 ∑ n = n 1 n 2 ⁢ ( e ⁡ [ n ] ) = ∑ n = n 1 n 2 ⁢ ( s ⁡ [ n ] + ∑ i = 1 M ⁢ a i ⁢ s ⁡ [ n - i ] ) .

### Claim 24 (depends on 14)

A window optimization method, as claimed in claim 14, wherein, calculating
50 ∂ J ∂ w ⁡ [ n ]
comprises, evaluating an equation for each of the window samples within the synthesis window
51 ∂ J ∂ w ⁡ [ n ] = ∑ k = n 1 n 2 ⁢ 2 ⁢ ⁢ e ⁡ [ k ] ⁢ ∂ e ⁡ [ k ] ∂ w ⁡ [ n ] = ∑ k = n 1 n 2 ⁢ 2 ⁢ e ⁡ [ k ] ⁢ ( ∑ i = 1 M ⁢ s ⁡ [ k - i ] ⁢ ∂ a i ∂ w ⁡ [ n ] ) ;
and defining the gradient by an equation
52 ∇ J = [ ∂ J ∂ w ⁡ [ 0 ] ⁢ ⁢ ∂ J ∂ w ⁡ [ 1 ] ⁢ ⁢ … ⁢ ⁢ ∂ J ∂ w ⁡ [ N - 1 ] ] T .

### Claim 25

A method for optimizing a window in linear prediction analysis of a speech signal, comprising:
assuming a rectangular initial window sequence, wherein the rectangular initial window sequence is a window sequence, wherein the window sequence comprises a plurality of window samples and wherein the length of the window sequence is N;
determining a gradient of a prediction error energy of the speech signal, wherein the speech signal is windowed by the rectangular initial window sequence, using a Levinson-Durbin initialization routine comprising:
defining a time lag l, wherein l equals zero;
determining an initial autocorrelation value with respect to each window sample of the rectangular initial window R[l], for l=0;
determining a partial derivative of the initial autocorrelation value with respect to each window sample of the rectangular initial window sequence, wherein a partial derivative of the initial autocorrelation value with respect to each window sample of the initial window sequence is indicated by
53 ∂ R ⁡ [ I ] ∂ w ⁡ [ n ]
wherein 1=0, and wherein determining R[l] for l=0 comprises determining R[l] for l=0 according to known values for l=0; and
determining a prediction error energy and a partial derivative of the prediction error energy as a function of the initial autocorrelation value with respect to each window sample of the rectangular initial window, wherein each of the prediction error energies are indicated by Jo and each of the partial derivatives of the prediction error energy is indicated by
54 ∂ J I ∂ w ⁡ [ n ]
wherein l=0;
updating the window sequence to create a next window sequence by defining the next window sequence as a function of a step size parameter, wherein the next window sequence becomes the window sequence;
determining a gradient of a new prediction error energy of the speech signal, wherein the speech signal is windowed by the window sequence;
wherein determining a gradient of a new prediction error energy of the speech signal comprises using a Levinson-Durbin recursion routine, wherein using a Levinson-Durbin recursion routine comprises:
determining a linear predictive coefficient and a partial derivative of the linear predictive coefficients for each of the window samples of the window sequence, wherein each of the linear predictive coefficients is indicated by an index i as ai and each of the partial derivatives of the linear predictive coefficients are indicated by
55 ∂ a i ∂ w ⁡ [ n ] ,
wherein determining the linear predictive coefficient and the partial derivative of the linear predictive coefficients for each of the window samples of the window sequence comprises using a Levinson-Durbin algorithm, wherein using a Levinson-Durbin algorithm comprises:
incrementing the time lag l, by defining l according to an equation l=I+1;
determining an l-order autocorrelation value with respect to each of the plurality of window samples of the window, wherein each of the l-order autocorrelation values is indicated by R[l], wherein determining R[I] comprises determining R[l] as a function of a plurality of indices k, the window length N, the plurality of speech signal samples s[k], and the plurality of window samples w[k] of the window sequence, wherein R[l] is defined by an equation
56 R ⁡ [ I ] = ∑ k = I N - 1 ⁢ w ⁡ [ k ] ⁢ s ⁡ [ k ] ⁢ w ⁡ [ k - I ] ⁢ s ⁡ [ k - I ] ;
determining a partial derivative of each of the l-order autocorrelation values with respect to each of the window samples of the window sequence, wherein each of the l-order autocorrelation values is indicated by
57 ∂ R ⁡ [ I ] ∂ w ⁡ [ n ] ,
wherein determining
58 ∂ R ⁡ [ I ] ∂ w ⁡ [ n ]
comprises determining
59 ∂ R ⁡ [ I ] ∂ w ⁡ [ n ]
according to known values;
calculating the linear predictive coefficients and the partial derivative of each of the linear predictive coefficients with respect to each of the window samples of the window sequence, wherein each of the linear predictive coefficients are indicated by an index i as ai and each of the partial derivatives of the linear predictive coefficients are indicated by
60 ∂ a i ∂ w ⁡ [ n ] ,
wherein calculating ai and
61 ∂ a i ∂ w ⁡ [ n ]
comprises:
determining a reflection coefficient for each of the window samples of the window sequences and a partial derivative of each of the reflection coefficients for each of the window samples of the window sequences, wherein each of the reflection coefficients are indicated by k, and the partial derivative of each of the reflection coefficients is indicated by
62 ∂ k l ∂ w ⁡ [ n ] ;
determining at least two update functions for each window sample of the window sequence and a partial derivative of each of the at least two update functions for each window sample of the window sequence, wherein the at least two update functions are indicated by a(l)i=−kl and a(l)i=a(l-1)l-1 and the partial derivative of each of the at least two update functions is indicated
63 ∂ a i ( l ) ∂ w ⁡ [ n ] = - ∂ k l ∂ w ⁡ [ n ] ⁢ ⁢ and ⁢ ∂ a i ( l ) ∂ w ⁡ [ n ] = ∂ a i ( l - 1 ) ∂ w ⁡ [ n ] - a l - i ( l - 1 ) ⁢ ∂ k l ∂ w ⁡ [ n ] - k l ⁢ ∂ a l - i ( l - 1 ) ∂ w ⁡ [ n ] ;
determining an l-order partial derivative of the linear predictive coefficients with respect to each window sample of the window sequence; and
determining if l equals M, wherein if l does not equal M, updating the l-order prediction error energy and the partial derivative of the prediction error energy, wherein the prediction error energy is indicated by J, and the partial derivative of the prediction error energy is indicated by
64 ∂ J l ∂ w ⁡ [ n ]
and repeating determining the at least two update functions and the partial derivative of each of the at least two update functions, for each window sample of the window sequence and determining if l equals M until l equals M; wherein when l equals M, defining the linear predictive coefficients according to an equation al=a(M)l and defining the partial derivative of the linear predictive coefficients according to an equation
65 ∂ a i ∂ w ⁡ [ n ] = ∂ a i ( M ) ∂ w ⁡ [ n ]
for each window sample of the window sequence;
determining if l equals an order M, wherein if l does not equal the order M, repeating the steps of incrementing the time lag l by defining l according to and equation l=l+1; determining R[l]; determining
66 ∂ R ⁡ [ l ] ∂ w ⁡ [ n ]
calculating the linear predictive coefficients and the partial derivatives of the linear predictive coefficients with respect to each of the window samples of the window sequence; and determining if l equals an order M until l equals an order M;
determining a prediction error sequence as a function of the speech signal windowed by the window sequence and the linear predictive coefficients, wherein the prediction error sequence comprises a new prediction energy estimate for each of the window samples of the window sequence, wherein determining the prediction error sequence as a function of the speech signal windowed by the window sequence and the linear predictive coefficients, comprises:
determining the prediction error sequence e[n] over a synthesis interval n wherein n ε[n1, n2], as defined by an equation,
67 ∑ n = n 1 n 2 ⁢ ⁢ ( ⅇ ⁡ [ n ] ) = ∑ n = n 1 n 2 ⁢ ⁢ ( s ⁡ [ n ] + ∑ i = 1 M ⁢ a i ⁢ s ⁡ [ n - i ] ) ;
determining a partial derivative of the new prediction energy estimate with respect to each of the window samples of the window sequence, wherein the partial derivative of the new prediction energy estimate with respect to each of the window samples of the window sequence is indicated by
68 ∂ J ∂ w ⁡ [ n ] ,
wherein, calculating
69 ∂ J ∂ w ⁡ [ n ]
comprises, evaluating an equation for each of the window samples within the synthesis window
70 ∂ J ∂ w ⁡ [ n ] = ∑ k = n 1 n 2 ⁢ ⁢ 2 ⁢ ⅇ ⁡ [ k ] ⁢ ∂ ⅇ ⁡ [ k ] ∂ w ⁡ [ n ] = ∑ k = n 1 n 2 ⁢ ⁢ 2 ⁢ ⅇ ⁡ [ k ] ⁢ ( ∑ i = 1 M ⁢ ⁢ s ⁡ [ k - i ] ⁢ ∂ a i ∂ w ⁡ [ n ] ) ;
and defining the gradient by an equation
71 ∇ J = [ ∂ J ∂ w ⁡ [ 0 ] ⁢ ⁢ ∂ J ∂ w ⁡ [ 1 ] ⁢ ⁢ ⋯ ⁢ ⁢ ∂ J ∂ w ⁡ [ N - 1 ] ] ;
and
determining whether a threshold has been reached; wherein if the threshold has not been reached, repeating the steps of updating the window to create the next window sequence, determining the gradient of the prediction error energy of the speech signal windowed by the window sequence wherein the next window sequence becomes the window sequence, and determining whether the threshold has been reached, until the threshold is reached.

### Claim 26

A method for optimizing a window in linear prediction analysis of a speech signal, comprising:
assuming an initial window sequence, wherein the initial window sequence is a window sequence, wherein the initial window sequence comprises a plurality of window samples, wherein each of the plurality of window samples of the initial window sequence is indicated by w[n], and wherein the length of the window sequence is N;
determining a prediction error energy as a function of the speech signal windowed by the initial window sequence;
updating the window sequence comprising, creating a perturbed window sequence as a function of a window perturbation constant, wherein the perturbed window sequence becomes the window sequence and the window sequence comprises a plurality of window samples, wherein each of the plurality of window samples of the perturbed window sequence is indicated by w′[n];
determining a new prediction error energy as a function of the speech signal windowed by the perturbed window sequence;
estimating a gradient of the new prediction error energy as a function of the speech signal windowed by the perturbed window sequence; and
determining whether a threshold has been reached; wherein if the threshold has not been reached, repeating the steps of updating the window sequence comprising, creating the next window sequence as the function of the window perturbation constant, wherein the perturbed window sequence becomes the window sequence; determining the new prediction error energy as the function of the speech signal windowed by the window sequence;
estimating the gradient of the prediction error energy as the function of the speech signal windowed by the window sequence, and determining whether the threshold has been reached, until the threshold is reached.

### Claim 27 (depends on 26)

A window optimization method, as claimed in claim 26, wherein assuming the initial window sequence comprises assuming a rectangular window sequence.

### Claim 28 (depends on 26)

A window optimization method, as claimed in claim 26, wherein determining the prediction error energy as the function of the speech signal windowed by the initial window sequence comprises using an autocorrelation method.

### Claim 29 (depends on 26)

A window optimization method, as claimed in claim 26, wherein creating the perturbed window sequence as the function of the window perturbation constant, wherein the window perturbation constant is indicated by Δw, comprises defining the perturbed window sequence according to a set of relationships comprising, w′[n]=w[n], n≠no; w′[no]=w[no]+Δw.

### Claim 30 (depends on 29)

A window optimization method, as claimed in claim 29, wherein the window perturbation constant has a value of approximately 10−7 to approximately 10−4.

### Claim 31 (depends on 26)

A window optimization method, as claimed in claim 26, wherein determining the new prediction error as a function of the speech signal windowed by the perturbed window sequence comprises, using an autocorrelation method.

### Claim 32 (depends on 31)

A window optimization method, as claimed in claim 31, wherein using the autocorrelation method comprises relating the new prediction error energy, wherein the new prediction error energy is indicated by J′[no], to perturbed autocorrelation values, wherein the perturbed autocorrelation values are indicated by R′[l,no], are a function of a time-lag l and sample no, according to a first equation J′[no]=R′[0,no]=R[0]+Δw (2w[no]+Δw)s2[no] for l=0 to a prediction order M and no=0 to N−1, and according to a second equation J′[no]=R′[l,no]=R[l]+Δw (w[no−l]s[no−l]+w[no+l]s[no+l])s[no] for l=0 to M and no=0 to N−1.

### Claim 33 (depends on 26)

A window optimization method, as claimed in claim 26, wherein estimating the gradient of the new prediction error energy as a function of the speech signal and the perturbed window sequence comprises, estimating the partial derivative of the new prediction error energy with respect to the window sequence for each of the window samples w′[no], wherein the partial derivative of the new prediction error energy with respect to the window sequence for each of the window samples is indicated by ∂J′l∂w[no].

### Claim 34 (depends on 33)

A window optimization method, as claimed in claim 33, wherein estimating the partial derivative of the new prediction error energy ∂J′l∂w[no] comprises, using an estimate based on a basic definition of a partial derivative.

### Claim 35 (depends on 34)

A window optimization method, as claimed in claim 34, wherein the basic definition of a derivative is defined by a function f(x), a variable x, an incremental change in the variable Δx, and by a relationship:
72 ∂ f ⁡ ( x ) ∂ x = lim Δ ⁢ ⁢ x → 0 ⁢ f ⁡ ( Δ ⁢ ⁢ x + x ) - f ⁡ ( x ) Δ ⁢ ⁢ x .

### Claim 36 (depends on 33)

A window optimization method, as claimed in claim 33, wherein estimating the partial derivative of the new prediction error energy, wherein the partial derivative of the new prediction error energy is indicated by ∂J′l∂w[no], comprises, defining the partial derivative of the prediction error energy for each window sample of the window sequence according to an equation (J′[no]−J)lΔw.

### Claim 37

A method for optimizing a window in linear prediction analysis of a speech signal, comprising:
assuming a rectangular initial window sequence, wherein the rectangular initial window sequence is a window sequence, wherein the rectangular initial window sequence comprises a plurality of window samples, wherein each of the plurality of window samples of the initial window sequence is indicated by w[n], and wherein the length of the window sequence is N;
determining a prediction error energy as a function of the speech signal windowed by the initial window sequence using an autocorrelation method;
updating the window sequence comprising, creating a perturbed window sequence as a function of a window perturbation constant, wherein the perturbed window sequence becomes the window sequence and the window sequence comprises a plurality of window samples, wherein each of the plurality of window samples of the perturbed window sequence is indicated by w′[n], and wherein creating the perturbed window sequence as the function of the window perturbation constant, wherein the window perturbation constant is indicated by Δw, comprises defining the perturbed window sequence according to a set of relationships comprising, w′[n]=w[n], n≠no; w′[no]=w[no]+Δw;
determining a new prediction error energy as a function of the speech signal windowed by the perturbed window sequence using an autocorrelation method, wherein using the autocorrelation method comprises relating the new prediction error energy, wherein the new prediction error energy is indicated by J′[no], to perturbed autocorrelation values, wherein the perturbed autocorrelation values are indicated by R′[l,no], are a function of a time-lag l and sample no, according to a first equation
J′[no]=R′[0,no]=R[0]+Δw(2w[no]+Δw)s2[no] for l=0 to a prediction order M and n0=0 to N−1, and according to a second equation J′[no]=R′[l,no]=R[l]+Δw(w[no−l]s[no−l]+w[no+l]s[no+l])s[no] for l=0 to M and no=0 to N−1;
estimating a gradient of the new prediction error energy as a function of the speech signal windowed by the perturbed window sequence comprising, estimating the partial derivative of the new prediction error energy with respect to the window sequence for each of the window samples w′[no], wherein the partial derivative of the new prediction error energy is indicated by ∂J′l∂w[no], comprises, defining the partial derivative of the prediction error energy for each window sample of the window sequence according to an equation (J′[no]−J)lΔw; and
determining whether a threshold has been reached; wherein if the threshold has not been reached, repeating the steps of updating the window sequence comprising, creating the next window sequence as the function of the window perturbation constant, wherein the perturbed window sequence becomes the window sequence; determining the new prediction error energy as the function of the speech signal windowed by the window sequence;
estimating the gradient of the prediction error energy as the function of the speech signal windowed by the window sequence, and determining whether the threshold has been reached, until the threshold is reached.

### Claim 38

A computer readable storage medium storing computer readable program code for producing an optimized window for analysis of a speech signal, the computer readable program code comprising:
data encoding the speech signal;
a computer code implementing a primary gradient-descent based window optimization procedure in response to an input of an initial window, wherein the primary gradient-descent based window optimization procedure optimizes the initial widow so as to minimize a prediction error energy by calculating a gradient of the prediction error energy.

### Claim 39

A computer readable storage medium storing computer readable program code for producing an optimized window for analysis of a speech signal, the computer readable program code comprising:
data encoding the speech signal;
a computer code implementing a primary gradient-descent based window optimization procedure in response to an input of an initial window, wherein the primary gradient-descent based window optimization procedure optimizes the initial widow so as to maximize a segmental prediction gain by calculating a gradient of a segmental prediction gain.

### Claim 40

A computer readable storage medium storing computer readable program code for producing an optimized window for analysis of a speech signal, the computer readable program code comprising:
data encoding the speech signal;
a computer code implementing an alternate gradient-descent based window optimization procedure in response to an input of an initial window, wherein the alternate gradient-descent based window optimization procedure optimizes the initial widow so as to minimize a prediction error energy by estimating a gradient of the prediction error energy.

### Claim 41

A computer readable storage medium storing computer readable program code for producing an optimized window for analysis of a speech signal, the computer readable program code comprising:
data encoding the speech signal;
a computer code implementing an alternate gradient-descent based window optimization procedure in response to an input of an initial window, wherein the alternate gradient-descent based window optimization procedure optimizes the initial widow so as to maximize a segmental prediction gain by estimating a gradient of a segmental prediction gain.

### Claim 42

A window optimization device, comprising:
a memory device, wherein the memory device stores a speech signal, at least one gradient-descent based window optimization procedure and know derivatives of autocorrelation values;
a processor coupled to the memory device, wherein the processor optimizes a window for linear predictive analysis of the speech signal using the speech signal, the at least one window optimization procedure and the known derivatives of the autocorrelation values communicated by the memory device.

### Claim 43 (depends on 42)

A window optimization device, as claimed in claim 42, wherein the at least one window gradient-descent based optimization procedure comprises a primary optimization procedure.

### Claim 44 (depends on 43)

A window optimization device, as claimed in claim 43, wherein the primary optimization procedure determines a gradient of a prediction error energy using a Levinson-Durbin based algorithm, wherein the Levinson-Durbin based algorithm is stored in the memory device and communicated to the processor.

### Claim 45 (depends on 42)

A window optimization device, as claimed in claim 42, wherein the at least one window gradient-descent based optimization procedure comprises an alternate optimization procedure.

### Claim 46 (depends on 45)

A window optimization device, as claimed in claim 45, wherein the alternate optimization procedure determines a gradient of a prediction error energy using an estimate based on a basic definition of a partial derivative, wherein the estimate based on a basic definition of a partial derivative is stored in the memory device and communicated to the processor.
