98
More On Linear Predictive Analysis 主主主 主主主

More On Linear Predictive Analysis 主講人:虞台文. Contents Linear Prediction Error Computation of the Gain Frequency Domain Interpretation of LPC Representations

Embed Size (px)

Citation preview

Page 1: More On Linear Predictive Analysis 主講人:虞台文. Contents Linear Prediction Error Computation of the Gain Frequency Domain Interpretation of LPC Representations

More On Linear Predictive Analysis

主講人:虞台文

Page 2: More On Linear Predictive Analysis 主講人:虞台文. Contents Linear Prediction Error Computation of the Gain Frequency Domain Interpretation of LPC Representations

Contents Linear Prediction Error Computation of the Gain Frequency Domain Interpretation of LPC Representations of LPC Coefficients

– Direct Representation– Roots of Predictor Polynomials– PARCO Coefficients– Log Area Ratio Coefficients– Line Spectrum Pair

Page 3: More On Linear Predictive Analysis 主講人:虞台文. Contents Linear Prediction Error Computation of the Gain Frequency Domain Interpretation of LPC Representations

More On Linear Predictive Analysis

Linear Prediction Error

Page 4: More On Linear Predictive Analysis 主講人:虞台文. Contents Linear Prediction Error Computation of the Gain Frequency Domain Interpretation of LPC Representations

LPC Error

1 ),()()( 00

αnGuknsαnep

kk

1 ),()()( 00

αnGuknsαnep

kk

Page 5: More On Linear Predictive Analysis 主講人:虞台文. Contents Linear Prediction Error Computation of the Gain Frequency Domain Interpretation of LPC Representations

Examples

PremphasizedSpeech Signals

Could be used for Pitch Detection

Page 6: More On Linear Predictive Analysis 主講人:虞台文. Contents Linear Prediction Error Computation of the Gain Frequency Domain Interpretation of LPC Representations

Normalized Mean-Squared Error

1

0

2

1

0

2

)(

)(

N

mn

pN

mn

n

ms

meV

1

0

2

1

0

2

)(

)(

N

mn

pN

mn

n

ms

meV

m

mn

ms

meV

)(

)(

2

2

m

mn

ms

meV

)(

)(

2

2

AutocorrelationMethod

1

0

2

1

0

2

)(

)(

N

mn

N

mn

n

ms

meV

1

0

2

1

0

2

)(

)(

N

mn

N

mn

n

ms

meVCovariance

Method

General Form

Page 7: More On Linear Predictive Analysis 主講人:虞台文. Contents Linear Prediction Error Computation of the Gain Frequency Domain Interpretation of LPC Representations

Normalized Mean-Squared Error

p

i

p

jj

ijin α

c

cαV

0 0 00

p

i

p

jj

ijin α

c

cαV

0 0 00

p

i

iin r

rαV

0 0

p

i

iin r

rαV

0 0

n

iin kV

1

2 )1(

n

iin kV

1

2 )1(

1

0

2

1

0

2

)(

)(

N

mn

pN

mn

n

ms

meV

1

0

2

1

0

2

)(

)(

N

mn

pN

mn

n

ms

meV

AutocorrelationMethod

1

0

2

1

0

2

)(

)(

N

mn

N

mn

n

ms

meV

1

0

2

1

0

2

)(

)(

N

mn

N

mn

n

ms

meVCovariance

Method

Page 8: More On Linear Predictive Analysis 主講人:虞台文. Contents Linear Prediction Error Computation of the Gain Frequency Domain Interpretation of LPC Representations

Experimental Evaluation on LPC Parameters

Frame Width N

Filter Order p

Conditions:

1. Covariance method and autocorrelation method

2. Synthetic vowel and Nature speech

3. Pitch synchronous and pitch asynchronous analysis

Page 9: More On Linear Predictive Analysis 主講人:虞台文. Contents Linear Prediction Error Computation of the Gain Frequency Domain Interpretation of LPC Representations

Pitch Synchronous Analysis

/i/

zero: the same order as the synthesizer.

The frame was beginning at the beginning of a pitch period.

Covariance method is more suitable for pitch synchronous analysis.

Covariance method is more suitable for pitch synchronous analysis.

Why the error increases?

Page 10: More On Linear Predictive Analysis 主講人:虞台文. Contents Linear Prediction Error Computation of the Gain Frequency Domain Interpretation of LPC Representations

Pitch Asynchronous Analysis

/i/

Both covariance and autocorrelation methods exhibit similar performance.

Both covariance and autocorrelation methods exhibit similar performance.

Monotonically decreasing

Page 11: More On Linear Predictive Analysis 主講人:虞台文. Contents Linear Prediction Error Computation of the Gain Frequency Domain Interpretation of LPC Representations

Frame Width Variation

/i/

Why the errors jump high when the frame size nears the multiples of pitch period?

Why the errors jump high when the frame size nears the multiples of pitch period?

The errors resulted by covariance and autocorrelation methods are compatible when N > 2P.

The errors resulted by covariance and autocorrelation methods are compatible when N > 2P.

Page 12: More On Linear Predictive Analysis 主講人:虞台文. Contents Linear Prediction Error Computation of the Gain Frequency Domain Interpretation of LPC Representations

Pitch Synchronous Analysis

Both for synthetic and nature speeches, covariance method is more suitable for pitch synchronous analysis.

Both for synthetic and nature speeches, covariance method is more suitable for pitch synchronous analysis.

Page 13: More On Linear Predictive Analysis 主講人:虞台文. Contents Linear Prediction Error Computation of the Gain Frequency Domain Interpretation of LPC Representations

Pitch Asynchronous Analysis

Both for synthetic and nature speeches, two methods are compatible.

Both for synthetic and nature speeches, two methods are compatible.

Page 14: More On Linear Predictive Analysis 主講人:虞台文. Contents Linear Prediction Error Computation of the Gain Frequency Domain Interpretation of LPC Representations

Frame Width Variation

Both for synthetic and nature speeches, the errors resulted by covariance and autocorrelation methods are compatible when N > 2P.

Both for synthetic and nature speeches, the errors resulted by covariance and autocorrelation methods are compatible when N > 2P.

Page 15: More On Linear Predictive Analysis 主講人:虞台文. Contents Linear Prediction Error Computation of the Gain Frequency Domain Interpretation of LPC Representations

More On Linear Predictive Analysis

Computation of the Gain

Page 16: More On Linear Predictive Analysis 主講人:虞台文. Contents Linear Prediction Error Computation of the Gain Frequency Domain Interpretation of LPC Representations

Speech Production Model (Review)

Impulse Train Generator

Impulse Train Generator

Random NoiseGenerator

Random NoiseGenerator

Time-VaryingDigital Filter

Time-VaryingDigital Filter

Vocal TractParameters

G

u(n)

s(n)

Page 17: More On Linear Predictive Analysis 主講人:虞台文. Contents Linear Prediction Error Computation of the Gain Frequency Domain Interpretation of LPC Representations

H(z)

Speech Production Model (Review)

Impulse Train Generator

Impulse Train Generator

Random NoiseGenerator

Random NoiseGenerator

Time-VaryingDigital Filter

Time-VaryingDigital Filter

Vocal TractParameters

G

u(n)

s(n)

p

k

kk za

1

1

1

p

k

kk za

1

1

1

p

k

kk za

G

zU

zSzH

1

1)(

)()(

p

k

kk za

G

zU

zSzH

1

1)(

)()(

)()()(1

nGuknsansp

kk

)()()(1

nGuknsansp

kk

Page 18: More On Linear Predictive Analysis 主講人:虞台文. Contents Linear Prediction Error Computation of the Gain Frequency Domain Interpretation of LPC Representations

Linear Prediction Model (Review)

)()()(1

nGuknsansp

kk

)()()(1

nGuknsansp

kk

p

kk knsαns

1

)()(ˆLinear Prediction:

)()(ˆ)( nensns Error compensation:

)()()(1

neknsαnsp

kk

)()()(1

neknsαnsp

kk

Page 19: More On Linear Predictive Analysis 主講人:虞台文. Contents Linear Prediction Error Computation of the Gain Frequency Domain Interpretation of LPC Representations

Speech Production vs. Linear Prediction

)()()(1

nGuknsansp

kk

)()()(1

nGuknsansp

kk

)()()(1

neknsαnsp

kk

)()()(1

neknsαnsp

kk

Speech production:

Linear Prediction:

Vocal Tract Excitation

Linear Predictor Error

ak = k

)()( nGune )()( nGune

Page 20: More On Linear Predictive Analysis 主講人:虞台文. Contents Linear Prediction Error Computation of the Gain Frequency Domain Interpretation of LPC Representations

Speech Production vs. Linear Prediction

)()()(1

nGuknsansp

kk

)()()(1

nGuknsansp

kk

)()()(1

neknsαnsp

kk

)()()(1

neknsαnsp

kk

Speech production:

Linear Prediction:

p

kk knsansnGu

1

)()()(

p

kk knsαnsne

1

)()()(

)()( nGune )()( nGune

Page 21: More On Linear Predictive Analysis 主講人:虞台文. Contents Linear Prediction Error Computation of the Gain Frequency Domain Interpretation of LPC Representations

The Gain

p

kk knsansnGu

1

)()()(

p

kk knsαnsne

1

)()()(

Generally, it is not possible to solve for G in a reliable way directly from the error signal itself.

Instead, we assume

1

0

221

0

2 )()(N

m

N

mn muGmeE

Energy of Error Energy of Excitation

Page 22: More On Linear Predictive Analysis 主講人:虞台文. Contents Linear Prediction Error Computation of the Gain Frequency Domain Interpretation of LPC Representations

1/A(z)

Assumptions about u(n)

1

0

221

0

2 )()(N

m

N

mn muGmeE

1

0

221

0

2 )()(N

m

N

mn muGmeE

Voiced Speech

Unvoiced Speech

)()( nδnu )()( nδnu

)()]()([ mδmnunuE )()]()([ mδmnunuE

This requires that both glottal pulse shape and lip radiation are lumped into the vocal tract model.

G(z)G(z) V(z)V(z) R(z)R(z)

G

u(n)=(n)h(n)

Page 23: More On Linear Predictive Analysis 主講人:虞台文. Contents Linear Prediction Error Computation of the Gain Frequency Domain Interpretation of LPC Representations

1/A(z)

Gain Estimation for Voiced Speech

1

0

221

0

2 )()(N

m

N

mn muGmeE

1

0

221

0

2 )()(N

m

N

mn muGmeE

Voiced Speech )()( nδnu )()( nδnu

This requires that both glottal pulse shape and lip radiation are lumped into the vocal tract model.

G(z)G(z) V(z)V(z) R(z)R(z)

G

u(n)=(n)h(n)

This requires that p is sufficiently large.This requires that p is sufficiently large.

Page 24: More On Linear Predictive Analysis 主講人:虞台文. Contents Linear Prediction Error Computation of the Gain Frequency Domain Interpretation of LPC Representations

1/A(z)

Gain Estimation for Voiced Speech

G(z)G(z) V(z)V(z) R(z)R(z)

G

u(n)=(n)h(n)

)()(

zA

GzH

)()(

zA

GzH (n) h(n)

p

k

kk zazA

1

1)(

p

k

kk zazA

1

1)(

Page 25: More On Linear Predictive Analysis 主講人:虞台文. Contents Linear Prediction Error Computation of the Gain Frequency Domain Interpretation of LPC Representations

Correlation Matching

)()(

zA

GzH

)()(

zA

GzH (n) h(n)

p

k

kk zazA

1

1)(

p

k

kk zazA

1

1)(

Define

n

ij jnhinhρ )()(

0

|)|()(n

jinhnh Assumed causal.

||̂ jir Autocorrelation function of the impulse response.

0

)()(ˆn

m nmhnhr

0

)()(ˆn

m nmhnhr

Page 26: More On Linear Predictive Analysis 主講人:虞台文. Contents Linear Prediction Error Computation of the Gain Frequency Domain Interpretation of LPC Representations

Correlation Matching

)()(

zA

GzH

)()(

zA

GzH (n) h(n)

p

k

kk zazA

1

1)(

p

k

kk zazA

1

1)(

0

)()(ˆn

m nmhnhr

0

)()(ˆn

m nmhnhr

Autocorrelation function of the speech signal

0

)()(n

m nmsnsr

0

)()(n

m nmsnsrIf H(z) correctly model the speech production system, we should have

mm rr ˆ mm rr ˆ

Page 27: More On Linear Predictive Analysis 主講人:虞台文. Contents Linear Prediction Error Computation of the Gain Frequency Domain Interpretation of LPC Representations

Correlation Matching

)()(

zA

GzH

)()(

zA

GzH (n) h(n)

p

k

kk zazA

1

1)(

p

k

kk zazA

1

1)(

0

)()(ˆn

m nmhnhr

0

)()(ˆn

m nmhnhr

0

)()(n

m nmsnsr

0

)()(n

m nmsnsr

GzAzH )()( )()(0

nGδinhap

ii

nn

p

ii jnhnGδjnhinha )()()()(

0

)()()(0

jGhjnhinhap

i ni

Page 28: More On Linear Predictive Analysis 主講人:虞台文. Contents Linear Prediction Error Computation of the Gain Frequency Domain Interpretation of LPC Representations

Correlation Matching

0

)()(ˆn

m nmhnhr

0

)()(ˆn

m nmhnhr

0

)()(n

m nmsnsr

0

)()(n

m nmsnsr

)()()(0

jGhjnhinhap

i ni

)(ˆ0

|| jGhrap

ijii

Assumed causal.

mm rr ˆ mm rr ˆ

2

0

Grap

iii

0j

00

||

p

ijiira pj ,,2,1

Page 29: More On Linear Predictive Analysis 主講人:虞台文. Contents Linear Prediction Error Computation of the Gain Frequency Domain Interpretation of LPC Representations

Correlation Matching

2

0

Grap

iii

0j

00

||

p

ijiira pj ,,2,1

The same formulation as autocorrelation method.

Page 30: More On Linear Predictive Analysis 主講人:虞台文. Contents Linear Prediction Error Computation of the Gain Frequency Domain Interpretation of LPC Representations

The Gain for voice speech

2

0

Grap

iii

0j

E n

nEG 2nEG 2

Page 31: More On Linear Predictive Analysis 主講人:虞台文. Contents Linear Prediction Error Computation of the Gain Frequency Domain Interpretation of LPC Representations

More on Autocorrelation

H(z)H(z)x(n) y(n))()()( knhkxny

k

Define )()(),( mnxnxmnxx

The stationary assumption implies

)]()([)],([)( mnxnxEmnEm xxxx )]()([)],([)( mnxnxEmnEm xxxx

AssumedStationary

Page 32: More On Linear Predictive Analysis 主講人:虞台文. Contents Linear Prediction Error Computation of the Gain Frequency Domain Interpretation of LPC Representations

Properties of LTI Systems

H(z)H(z)x(n) y(n)( ) ( ) ( )

k

y n h k x n k

Define )()(),( mnynymnyy

)]()([)],([)( mnxnxEmnEm xxxx )]()([)],([)( mnxnxEmnEm xxxx

lk

lmnxlhknxkh )()()()(

lk

lmnxknxlhkh )()()()(

Page 33: More On Linear Predictive Analysis 主講人:虞台文. Contents Linear Prediction Error Computation of the Gain Frequency Domain Interpretation of LPC Representations

Properties of LTI Systems

)]()([)],([)( mnxnxEmnEm xxxx )]()([)],([)( mnxnxEmnEm xxxx

])()([)()()],([

lk

yy lmnxknxElhkhmnE

][)()(

l

xxk

lkmlhkh

Define )()(),( mnynymnyy

lk

lmnxlhknxkh )()()()(

lk

lmnxknxlhkh )()()()(

Page 34: More On Linear Predictive Analysis 主講人:虞台文. Contents Linear Prediction Error Computation of the Gain Frequency Domain Interpretation of LPC Representations

Properties of LTI Systems

)]()([)],([)( mnxnxEmnEm xxxx )]()([)],([)( mnxnxEmnEm xxxx

])()([)()()],([

lk

yy lmnxknxElhkhmnE

][)()(

l

xxk

lkmlhkh

Independent on n

)]()([)],([)( mnynyEmnEm yyyy )]()([)],([)( mnynyEmnEm yyyy

y(n) is also stationary.

Page 35: More On Linear Predictive Analysis 主講人:虞台文. Contents Linear Prediction Error Computation of the Gain Frequency Domain Interpretation of LPC Representations

Properties of LTI Systems

)]()([)],([)( mnxnxEmnEm xxxx )]()([)],([)( mnxnxEmnEm xxxx

][)()(][

l

xxk

yy lkmlhkhm

][)()(

k l

xx lkmlhkh ll+k

][)()(

k l

xx lmklhkh

l k

xx klhkhlm )()(][

Page 36: More On Linear Predictive Analysis 主講人:虞台文. Contents Linear Prediction Error Computation of the Gain Frequency Domain Interpretation of LPC Representations

Properties of LTI Systems

l k

xxyy klhkhlmm )()(][][

k

klhkhlc )()()(

k

klhkhlc )()()(Define

l

xxyy lclmm )(][][

Estimatedfrom input

Estimatedfrom output

FilterDesign

Page 37: More On Linear Predictive Analysis 主講人:虞台文. Contents Linear Prediction Error Computation of the Gain Frequency Domain Interpretation of LPC Representations

The Gain for Unvoiced Speech

)()]()([ mδmnunuE )()]()([ mδmnunuE

)()(

zA

GzH

)()(

zA

GzH u(n) s(n)

p

k

kk zazA

1

1)(

p

k

kk zazA

1

1)(

)()()(1

nGuknsansp

kk

)()()()]()([1

mnsnGuknsaEmnsnsEp

kk

p

kkm mnsnuGEmnsknsEar

1

)]()([)]()([ˆ

p

kkm mnsnuGEmnsknsEar

1

)]()([)]()([ˆ

Page 38: More On Linear Predictive Analysis 主講人:虞台文. Contents Linear Prediction Error Computation of the Gain Frequency Domain Interpretation of LPC Representations

The Gain for Unvoiced Speech

p

kkm mnsnuGEmnsknsEar

1

)]()([)]()([ˆ

p

kkm mnsnuGEmnsknsEar

1

)]()([)]()([ˆ

p

kkmkm mnsnuGErar

1|| )]()([ˆˆ

=?

Page 39: More On Linear Predictive Analysis 主講人:虞台文. Contents Linear Prediction Error Computation of the Gain Frequency Domain Interpretation of LPC Representations

The Gain for Unvoiced Speech

p

kkmkm mnsnuGErar

1|| )]()([ˆˆ

=?

0 ,0)]()([ mmnsnuE

Why?

:0m

)()()()]()([1

nGuknsanuEnsnuEp

kk

GnunuGE )()(

Page 40: More On Linear Predictive Analysis 主講人:虞台文. Contents Linear Prediction Error Computation of the Gain Frequency Domain Interpretation of LPC Representations

The Gain for Unvoiced Speech

p

kkmkm mnsnuGErar

1|| )]()([ˆˆ

0 ,0)]()([ mmnsnuE

0 ,)]()([ mGmnsnuE

Estimated using rm

pmrap

kkmk ,2,1 ,0

0||

0 ,0

2

mGrap

kkk

Page 41: More On Linear Predictive Analysis 主講人:虞台文. Contents Linear Prediction Error Computation of the Gain Frequency Domain Interpretation of LPC Representations

The Gain for Unvoiced Speech

pmrap

kkmk ,2,1 ,0

0||

0 ,0

2

mGrap

kkk

Once again, we have the same formulation as autocorrelation method.

Furthermore, nEG 2nEG 2

Page 42: More On Linear Predictive Analysis 主講人:虞台文. Contents Linear Prediction Error Computation of the Gain Frequency Domain Interpretation of LPC Representations

More On Linear Predictive Analysis

Frequency Domain Interpretation of LPC

Page 43: More On Linear Predictive Analysis 主講人:虞台文. Contents Linear Prediction Error Computation of the Gain Frequency Domain Interpretation of LPC Representations

Spectral Representation of Vocal Tract

)(1

)(

1

zA

G

za

GzH p

k

kk

)(1

)(

1

jp

k

kjk

j

eA

G

ea

GeH

Page 44: More On Linear Predictive Analysis 主講人:虞台文. Contents Linear Prediction Error Computation of the Gain Frequency Domain Interpretation of LPC Representations

Spectra

Page 45: More On Linear Predictive Analysis 主講人:虞台文. Contents Linear Prediction Error Computation of the Gain Frequency Domain Interpretation of LPC Representations

Frequency Domain Interpretation of Mean-Squared Prediction Error

)(1

0

2 meEpN

mnn

)()()( zAzSzE nn

Parseval’s Theorem

dωeAeSπ

E jωπ

π

jωnn

22 |)(||)(|2

1

Page 46: More On Linear Predictive Analysis 主講人:虞台文. Contents Linear Prediction Error Computation of the Gain Frequency Domain Interpretation of LPC Representations

Frequency Domain Interpretation of Mean-Squared Prediction Error

dωeAeSπ

E jωπ

π

jωnn

22 |)(||)(|2

1

)()(

jωjω

eA

GeH

)()(

jωjω

eH

GeA

dωeH

eS

πE

π

π jω

jωn

n 2

2

|)(|

|)(|

2

1 dωeH

eS

πE

π

π jω

jωn

n 2

2

|)(|

|)(|

2

1

Page 47: More On Linear Predictive Analysis 主講人:虞台文. Contents Linear Prediction Error Computation of the Gain Frequency Domain Interpretation of LPC Representations

Frequency Domain Interpretation of Mean-Squared Prediction Error

dωeH

eS

πE

π

π jω

jωn

n 2

2

|)(|

|)(|

2

1 dωeH

eS

πE

π

π jω

jωn

n 2

2

|)(|

|)(|

2

1

|Sn(ej)| > |H(ej)| contributes more to the total error than |Sn(ej)| < |H(ej)|.

Hence, the LPC spectral error criterion favors a good fit near the spectral peak.

Page 48: More On Linear Predictive Analysis 主講人:虞台文. Contents Linear Prediction Error Computation of the Gain Frequency Domain Interpretation of LPC Representations

Spectra

Page 49: More On Linear Predictive Analysis 主講人:虞台文. Contents Linear Prediction Error Computation of the Gain Frequency Domain Interpretation of LPC Representations

More On Linear Predictive Analysis

Representations of LPC Coefficients ---

Direct Representation

Page 50: More On Linear Predictive Analysis 主講人:虞台文. Contents Linear Prediction Error Computation of the Gain Frequency Domain Interpretation of LPC Representations

Direct Representation

pp zazazazA 2

21

11)(

Coding ai’s directly.

z1

z1

z1

a1

a2

ap

uG[n] uL[n]G

G/A(z)G/A(z)

Page 51: More On Linear Predictive Analysis 主講人:虞台文. Contents Linear Prediction Error Computation of the Gain Frequency Domain Interpretation of LPC Representations

Disadvantages

The dynamic ranges of ai’s is relatively large.

Quantatization possibly causes instability problems.

Page 52: More On Linear Predictive Analysis 主講人:虞台文. Contents Linear Prediction Error Computation of the Gain Frequency Domain Interpretation of LPC Representations

More On Linear Predictive Analysis

Representations of LPC Coefficients ---

Roots of Predictor Polynomials

Page 53: More On Linear Predictive Analysis 主講人:虞台文. Contents Linear Prediction Error Computation of the Gain Frequency Domain Interpretation of LPC Representations

Roots of the Predictor Polynomial

pp zazazazA 2

21

11)(

p

kk zz

1

1)1(

Coding p/2 zk’s.

kjkk erz

Dynamic range of rk’s?

Dynamic range of k’s?

10 kr

k0

Page 54: More On Linear Predictive Analysis 主講人:虞台文. Contents Linear Prediction Error Computation of the Gain Frequency Domain Interpretation of LPC Representations

The Application

2/

1

11 )1)(1()(p

k

jk

jk zerzerzA kk

2/

1

221cos21p

kkkk zrzr

kbk er Let kk rb ln

2/

1

221cos21p

k

bk

b zeze kk

Formant Analysis Application.Formant Analysis Application.

Page 55: More On Linear Predictive Analysis 主講人:虞台文. Contents Linear Prediction Error Computation of the Gain Frequency Domain Interpretation of LPC Representations

Implementation

Each Stage represents one formant frequency and its corresponding bandwidth.

11 cos2 r

21r

22 cos2 r

22r

2/2/ cos2 ppr

22/pr

][nuG ][nuLG

z1

z1

z1

z1

z1

z1

G/A(z)G/A(z)

Page 56: More On Linear Predictive Analysis 主講人:虞台文. Contents Linear Prediction Error Computation of the Gain Frequency Domain Interpretation of LPC Representations

More On Linear Predictive Analysis

Representations of LPC Coefficients ---

PARCO Coefficients

Page 57: More On Linear Predictive Analysis 主講人:虞台文. Contents Linear Prediction Error Computation of the Gain Frequency Domain Interpretation of LPC Representations

PARCO Coefficients

1)(0 na

1,,2,1 ,)1()1()(

niakaa ninn

ni

ni

nn

n ka )(

Step-Up Procedure:

piaa pii ,,2,0 )( piaa p

ii ,,2,0 )(

Page 58: More On Linear Predictive Analysis 主講人:虞台文. Contents Linear Prediction Error Computation of the Gain Frequency Domain Interpretation of LPC Representations

PARCO Coefficients

1,,2,1 ,1 2

)()()1(

nik

akaa

n

ninn

nin

i

)(nnn ak

Step-Down Procedure:

where n goes from p to p1, down to 1 and initially we set:

piaa ip

i ,,2,1 ,)(

Dynamic range of ki’s? 11 ik

Page 59: More On Linear Predictive Analysis 主講人:虞台文. Contents Linear Prediction Error Computation of the Gain Frequency Domain Interpretation of LPC Representations

More On Linear Predictive Analysis

Representations of LPC Coefficients ---

Log Area Ratio Coefficients

Page 60: More On Linear Predictive Analysis 主講人:虞台文. Contents Linear Prediction Error Computation of the Gain Frequency Domain Interpretation of LPC Representations

Log Area Ratio Coefficients

ki’s: Reflection Coefficients

ii

iii AA

AAk

1

1

1/

1/

1

1

ii

ii

AA

AA

i

i

i

i

k

k

A

A

1

11

gi’s: Log Area Ratios

1

1logi

ii A

Ag

i

i

k

k

1

1log

1

1

i

i

g

i g

ek

e

Page 61: More On Linear Predictive Analysis 主講人:虞台文. Contents Linear Prediction Error Computation of the Gain Frequency Domain Interpretation of LPC Representations

More On Linear Predictive Analysis

Representations of LPC Coefficients ---

Line Spectrum Pair

Page 62: More On Linear Predictive Analysis 主講人:虞台文. Contents Linear Prediction Error Computation of the Gain Frequency Domain Interpretation of LPC Representations

LPC Coefficients

mmm zazazazA 2

21

11)(

where m is the order of the inverse filter.

Line Spectrum Pair (LSP) is an alternative LPC spectral representation.

If the system is stable, all zeros of the inverse filter are inside the unit circle.

Page 63: More On Linear Predictive Analysis 主講人:虞台文. Contents Linear Prediction Error Computation of the Gain Frequency Domain Interpretation of LPC Representations

Line Spectrum Pair LSP contains two polynomials. The zeros of the the two polynomials have the f

ollowing properties:– Lie on unit circle– Interlaced– Through quantization, the minimum phase property

of the filter is kept.

Useful for vocoder application.

Page 64: More On Linear Predictive Analysis 主講人:虞台文. Contents Linear Prediction Error Computation of the Gain Frequency Domain Interpretation of LPC Representations

Recursive Relation of the inverse filter

)()()( 1)1(11

zAzkzAzA m

mmmm

where km+1 is the reflection coefficient of the m+1th tube.

Recall that

ii

iii AA

AAk

1

1

ii

iii AA

AAk

1

1ki =1:

ki =1:

Special cases:

Ai+1

Ai+1=0

Page 65: More On Linear Predictive Analysis 主講人:虞台文. Contents Linear Prediction Error Computation of the Gain Frequency Domain Interpretation of LPC Representations

LSP Polynomials

)()()( 1)1( zAzzAzP mm

m

)()()( 1)1( zAzzAzQ mm

m

)()()( 21 zQzPzAm )()()( 21 zQzPzAm

mmm zazazazA 2

21

11)(

Page 66: More On Linear Predictive Analysis 主講人:虞台文. Contents Linear Prediction Error Computation of the Gain Frequency Domain Interpretation of LPC Representations

Properties of LSP Polynomials

)()()( 21 zQzPzAm )()()( 21 zQzPzAm

Show thatThe zeros of P(z) and Q(z) are on the unit circle and interlaced.

Page 67: More On Linear Predictive Analysis 主講人:虞台文. Contents Linear Prediction Error Computation of the Gain Frequency Domain Interpretation of LPC Representations

Proof )()()( 1)1( zAzzAzP m

mm

)()()( 1)1( zAzzAzP mm

m

)()()( 1)1( zAzzAzQ mm

m)()()( 1)1( zAzzAzQ m

mm

)(

)(1)()(

1)1(

zA

zAzzAzP

m

mmm

)(

)(1)()(

1)1(

zA

zAzzAzQ

m

mmm

)(

)()(

1)1(

zA

zAzzH

m

mm

)(

)()(

1)1(

zA

zAzzH

m

mm

)(1)( zHzAm

)(1)( zHzAm

Page 68: More On Linear Predictive Analysis 主講人:虞台文. Contents Linear Prediction Error Computation of the Gain Frequency Domain Interpretation of LPC Representations

Proof

)(

)()(

1)1(

zA

zAzzH

m

mm

)(

)()(

1)1(

zA

zAzzH

m

mm

)(1)()( zHzAzP m )(1)()( zHzAzP m

)(1)()( zHzAzQ m )(1)()( zHzAzQ m

m

iim zzzA

1

1)1()(

m

iim zzzA

1

1 )1()(

m

iim

m zzzzAz1

111)1( )()(

m

i i

i

zz

zzz

11

11

)1(

)(

m

i i

i

zz

zzzzH

1

1

)(

)1()(

m

i i

i

zz

zzzzH

1

1

)(

)1()(

1 , ijω

ii rerz i 1 , ijω

ii rerz i

Page 69: More On Linear Predictive Analysis 主講人:虞台文. Contents Linear Prediction Error Computation of the Gain Frequency Domain Interpretation of LPC Representations

Proof

m

i i

i

zz

zzzzH

1

1

)(

)1()(

m

i i

i

zz

zzzzH

1

1

)(

)1()(

1 , ijω

ii rerz i 1 , ijω

ii rerz i

)||1)(||1(|||1| 2222iii zzzzzz )||1)(||1(|||1| 2222

iii zzzzzz

)1)(1(|1| **2 zzzzzz iii **22 ||||1 zzzzzz iii

))((|| **2iii zzzzzz

zzzzzz iii**22 ||||

)(1)()( zHzAzP m )(1)()( zHzAzP m

)(1)()( zHzAzQ m )(1)()( zHzAzQ m

Page 70: More On Linear Predictive Analysis 主講人:虞台文. Contents Linear Prediction Error Computation of the Gain Frequency Domain Interpretation of LPC Representations

Proof )||1)(||1(|||1| 2222

iii zzzzzz )||1)(||1(|||1| 2222iii zzzzzz

m

i i

i

zz

zzzzH

1

1

)(

)1()(

m

i i

i

zz

zzzzH

1

1

)(

)1()(

1 , ijω

ii rerz i 1 , ijω

ii rerz i

>0

1||1

1||1

1||1

|)(|

z

z

z

zH

)(1)()( zHzAzP m )(1)()( zHzAzP m

)(1)()( zHzAzQ m )(1)()( zHzAzQ m

Page 71: More On Linear Predictive Analysis 主講人:虞台文. Contents Linear Prediction Error Computation of the Gain Frequency Domain Interpretation of LPC Representations

Proof )||1)(||1(|||1| 2222

iii zzzzzz )||1)(||1(|||1| 2222iii zzzzzz

m

i i

i

zz

zzzzH

1

1

)(

)1()(

m

i i

i

zz

zzzzH

1

1

)(

)1()(

1 , ijω

ii rerz i 1 , ijω

ii rerz i

1||1

1||1

1||1

|)(|

z

z

z

zH

)(1)()( zHzAzP m )(1)()( zHzAzP m

)(1)()( zHzAzQ m )(1)()( zHzAzQ m

P(z)=0Q(z)=0

iff H(z) = 1.

This concludes that the zeros of P(z) and Q(z) are on the unit circle.

Page 72: More On Linear Predictive Analysis 主講人:虞台文. Contents Linear Prediction Error Computation of the Gain Frequency Domain Interpretation of LPC Representations

Proof (interlaced zeros)

m

i i

i

zz

zzzzH

1

1

)(

)1()(

m

i i

i

zz

zzzzH

1

1

)(

)1()( 1 , i

jωii rerz i 1 , i

jωii rerz i

Fact: H(z) is an all-pass filter.

i

iap zz

zzzH

i

*1

)(

)cos(1

)sin(tan2)( 1

ii

iii ωωr

ωωrωωφ

)()( ωjφjω eeH )()( ωjφjω eeH

)()( ωφjωap

i

ieeH

m

i ii

ii

ωωr

ωωrωmωφ

1

1

)cos(1

)sin(tan2)1()(

m

i ii

ii

ωωr

ωωrωmωφ

1

1

)cos(1

)sin(tan2)1()( Phase

One can verify that (0) = 0 and (2) = 2(m+1)

Page 73: More On Linear Predictive Analysis 主講人:虞台文. Contents Linear Prediction Error Computation of the Gain Frequency Domain Interpretation of LPC Representations

Proof (interlaced zeros)

)()( ωjφjω eeH )()( ωjφjω eeH

m

i ii

ii

ωωr

ωωrωmωφ

1

1

)cos(1

)sin(tan2)1()(

m

i ii

ii

ωωr

ωωrωmωφ

1

1

)cos(1

)sin(tan2)1()( Phase

πnωφ

nπωφeH jω

)12()(1

2)(1)(

)(1)()( zHzAzP m )(1)()( zHzAzP m

)(1)()( zHzAzQ m )(1)()( zHzAzQ m

Therefore, z=1 is a zero of Q(z).

zeros of P(z)

zeros of Q(z)

One can verify that (0) = 0 and (2) = 2(m+1)

Page 74: More On Linear Predictive Analysis 主講人:虞台文. Contents Linear Prediction Error Computation of the Gain Frequency Domain Interpretation of LPC Representations

Proof (interlaced zeros)

m

i ii

ii

ωωr

ωωrωmωφ

1

1

)cos(1

)sin(tan2)1()(

m

i ii

ii

ωωr

ωωrωmωφ

1

1

)cos(1

)sin(tan2)1()(

(0) = 0

(2) = 2(m+1)

0 2

()

2(m+1)

Is this possible?Is this possible?

Page 75: More On Linear Predictive Analysis 主講人:虞台文. Contents Linear Prediction Error Computation of the Gain Frequency Domain Interpretation of LPC Representations

Proof (interlaced zeros)

m

i ii

ii

ωωr

ωωrωmωφ

1

1

)cos(1

)sin(tan2)1()(

m

i ii

ii

ωωr

ωωrωmωφ

1

1

)cos(1

)sin(tan2)1()(

0 2

()

2(m+1)

Is this possible?Is this possible?

(0) = 0

(2) = 2(m+1)

Page 76: More On Linear Predictive Analysis 主講人:虞台文. Contents Linear Prediction Error Computation of the Gain Frequency Domain Interpretation of LPC Representations

Proof (interlaced zeros)

m

i ii

ii

ωωr

ωωrωmωφ

1

1

)cos(1

)sin(tan2)1()(

m

i ii

ii

ωωr

ωωrωmωφ

1

1

)cos(1

)sin(tan2)1()(

Group Delay

ω

ωφωτ

)(

)(

m

i iii

i

ωωrr

r

12

2

)cos(21

11 > 0

() is monotonically decreasing.

(0) = 0

(2) = 2(m+1)

Page 77: More On Linear Predictive Analysis 主講人:虞台文. Contents Linear Prediction Error Computation of the Gain Frequency Domain Interpretation of LPC Representations

Proof (interlaced zeros)

m

i ii

ii

ωωr

ωωrωmωφ

1

1

)cos(1

)sin(tan2)1()(

m

i ii

ii

ωωr

ωωrωmωφ

1

1

)cos(1

)sin(tan2)1()(

0 2

()

2(m+1)

2345 ...

Typical shape of ()

Typical shape of ()

(0) = 0

(2) = 2(m+1)

Page 78: More On Linear Predictive Analysis 主講人:虞台文. Contents Linear Prediction Error Computation of the Gain Frequency Domain Interpretation of LPC Representations

m

i ii

ii

ωωr

ωωrωmωφ

1

1

)cos(1

)sin(tan2)1()(

m

i ii

ii

ωωr

ωωrωmωφ

1

1

)cos(1

)sin(tan2)1()(

0 2

()

2(m+1)

2345 ...

Typical shape of ()

Typical shape of ()

Proof (interlaced zeros)

)(1)()( zHzAzP m )(1)()( zHzAzP m

)(1)()( zHzAzQ m )(1)()( zHzAzQ m

πnωφ

nπωφeH jω

)12()(1

2)(1)(

πnωφ

nπωφeH jω

)12()(1

2)(1)(

Q(ej)=0Q(ej)=0

P(ej)=0P(ej)=0

Q(ej)=0Q(ej)=0

P(ej)=0P(ej)=0

Q(ej)=0Q(ej)=0

P(ej)=0P(ej)=0

Page 79: More On Linear Predictive Analysis 主講人:虞台文. Contents Linear Prediction Error Computation of the Gain Frequency Domain Interpretation of LPC Representations

Proof (interlaced zeros) 0 2

()

2(m+1)

2345 ...

Typical shape of ()

Typical shape of ()

Q(ej)=0Q(ej)=0

P(ej)=0P(ej)=0

Q(ej)=0Q(ej)=0

P(ej)=0P(ej)=0

Q(ej)=0Q(ej)=0

P(ej)=0P(ej)=0

There are 2(m+1) cross points from 0 , these constitute the 2(m+1) interlaced zeros of P(z) and Q(z).

Page 80: More On Linear Predictive Analysis 主講人:虞台文. Contents Linear Prediction Error Computation of the Gain Frequency Domain Interpretation of LPC Representations

Quantization of LSP Zeros

)()()( 21 zQzPzAm )()()( 21 zQzPzAm

m

i

j zezP i

2

1

1)1()( 2

m

i

j zezQ i

2

1

1)1()( 12

For effective transmission, we quantize i’s into several levels, e.g., using 5 bits.

Is such a quantization detrimental? Is such a quantization detrimental?

Page 81: More On Linear Predictive Analysis 主講人:虞台文. Contents Linear Prediction Error Computation of the Gain Frequency Domain Interpretation of LPC Representations

Minimum Phase Preserving Property

Show that in quantizing the LSP frequencies, the reconstructed all-pole filter preserves its minimum phase property as long as the zeros has the properties shown in the left figure.

Page 82: More On Linear Predictive Analysis 主講人:虞台文. Contents Linear Prediction Error Computation of the Gain Frequency Domain Interpretation of LPC Representations

Find the Roots of P(z) and Q(z)

mmm zazazazA 2

21

11)(

)()()( 1)1( zAzzAzP mm

m)()()( 1)1( zAzzAzP m

mm

)()()( 1)1( zAzzAzQ mm

m)()()( 1)1( zAzzAzQ m

mm

mmm zazazazA 2

21

11 1)(

)1(1

21

11)1( )(

mmmmm

m zzazazazAz

)1(

11)(1)(

m

m

i

iimi zzaazP

)1(

11)(1)(

m

m

i

iimi zzaazQ

Symmetric

Anti-symmetric

Page 83: More On Linear Predictive Analysis 主講人:虞台文. Contents Linear Prediction Error Computation of the Gain Frequency Domain Interpretation of LPC Representations

Find the Roots of P(z) and Q(z)

)1(

11)(1)(

m

m

i

iimi zzaazP

)1(

11)(1)(

m

m

i

iimi zzaazP )1(

11)(1)(

m

m

i

iimi zzaazQ

)1(

11)(1)(

m

m

i

iimi zzaazQ

)1)(1()( 22

11

1 mm zpzpzpzzP

11 1 paa m

2112 ppaa m

3223 ppaa m

1)( 11 maap

1122 )( paap m

2233 )( paap m

/ 2 / 2 1 / 2 1 / 2m m m ma a p p / 2 / 2 / 2 1 / 2 1( )m m m mp a a p ...

...

/ 2 1 / 2 1m mp p ...

11 ppm

1mp

...

Page 84: More On Linear Predictive Analysis 主講人:虞台文. Contents Linear Prediction Error Computation of the Gain Frequency Domain Interpretation of LPC Representations

Find the Roots of P(z) and Q(z)

)1(

11)(1)(

m

m

i

iimi zzaazP

)1(

11)(1)(

m

m

i

iimi zzaazP )1(

11)(1)(

m

m

i

iimi zzaazQ

)1(

11)(1)(

m

m

i

iimi zzaazQ

1

)(

0

11

p

paap iimii

1

)(

0

11

p

paap iimii

We only need compute the values on 1 i m/2.

)1)(1()( 22

11

1 mm zpzpzpzzP

1)( 11 maap

1122 )( paap m

2233 )( paap m

/ 2 / 2 / 2 1 / 2 1( )m m m mp a a p ...

/ 2 1 / 2 1m mp p ...

11 ppm

1mp

Page 85: More On Linear Predictive Analysis 主講人:虞台文. Contents Linear Prediction Error Computation of the Gain Frequency Domain Interpretation of LPC Representations

Find the Roots of P(z) and Q(z)

)1(

11)(1)(

m

m

i

iimi zzaazP

)1(

11)(1)(

m

m

i

iimi zzaazP )1(

11)(1)(

m

m

i

iimi zzaazQ

)1(

11)(1)(

m

m

i

iimi zzaazQ

)1)(1()( 22

11

1 mm zqzqzqzzQ

11 1 qaa m

2112 qqaa m

3223 qqaa m

1)( 11 maaq

1122 )( qaaq m

2233 )( qaaq m

/ 2 / 2 1 / 2 1 / 2m m m ma a q q / 2 / 2 / 2 1 / 2 1( )m m m mq a a q ...

...

/ 2 1 / 2 1m mq q ...

11 qqm

1mq

...

Page 86: More On Linear Predictive Analysis 主講人:虞台文. Contents Linear Prediction Error Computation of the Gain Frequency Domain Interpretation of LPC Representations

Find the Roots of P(z) and Q(z)

)1(

11)(1)(

m

m

i

iimi zzaazP

)1(

11)(1)(

m

m

i

iimi zzaazP )1(

11)(1)(

m

m

i

iimi zzaazQ

)1(

11)(1)(

m

m

i

iimi zzaazQ

1

)(

0

11

q

qaaq iimii

1

)(

0

11

q

qaaq iimii

We only need compute the values on 1 i m/2.

)1)(1()( 22

11

1 mm zqzqzqzzQ

1)( 11 maaq

1122 )( qaaq m

2233 )( qaaq m

/ 2 / 2 / 2 1 / 2 1( )m m m mq a a q ...

/ 2 1 / 2 1m mq q ...

11 qqm

1mq

Page 87: More On Linear Predictive Analysis 主講人:虞台文. Contents Linear Prediction Error Computation of the Gain Frequency Domain Interpretation of LPC Representations

Find the Roots of P(z) and Q(z)

)1)(1()( )1(1

22

11

1 mm zzpzpzpzzP

)1)(1()( )1(1

22

11

1 mm zzqzqzqzzQ

zeroon 1

zeroon +1

P’(z)

Q’(z)

Both P’(z) and Q’(z) are symmetric. Both P’(z) and Q’(z) are symmetric.

Page 88: More On Linear Predictive Analysis 主講人:虞台文. Contents Linear Prediction Error Computation of the Gain Frequency Domain Interpretation of LPC Representations

Find the Roots of P(z) and Q(z)

mm zzpzpzpzP )1(1

22

111)(

jmmjjjj eepepepeP )1(1

2211)(

12/

12/

)()( 22222

m

im

ijiji

jjj peepeeemmmmm

2

/ 2 11

/ 22 2 21

2 cos( ) cos( )m

mj m m

i mi

e p i p

)(~ P To find its zeros.

Page 89: More On Linear Predictive Analysis 主講人:虞台文. Contents Linear Prediction Error Computation of the Gain Frequency Domain Interpretation of LPC Representations

Find the Roots of P(z) and Q(z)

/ 2 11

/ 22 2 21

( ) cos( ) cos( )m

m mi m

i

P p i p

/ 2 1

1/ 22 2 2

1

( ) cos( ) cos( )m

m mi m

i

P p i p

)cos()cos(coscos2 )cos()cos(coscos2

cos3cos43cos 3

)2(2cos4cos 1)2(cos2 2

cos5cos2cos3cos2

cos2cos3cos25cos

1cos22cos 2

Page 90: More On Linear Predictive Analysis 主講人:虞台文. Contents Linear Prediction Error Computation of the Gain Frequency Domain Interpretation of LPC Representations

Find the Roots of P(z) and Q(z)

Define kck cos xc cos1

odd is 2

even is 12

2/)1(2/)1(

22/

kxcc

kcc

kk

kk

odd is 2

even is 12

2/)1(2/)1(

22/

kxcc

kcc

kk

kk

)cos()cos(coscos2 )cos()cos(coscos2

/ 2 11

/ 22 2 21

( ) cos( ) cos( )m

m mi m

i

P p i p

/ 2 1

1/ 22 2 2

1

( ) cos( ) cos( )m

m mi m

i

P p i p

Page 91: More On Linear Predictive Analysis 主講人:虞台文. Contents Linear Prediction Error Computation of the Gain Frequency Domain Interpretation of LPC Representations

Find the Roots of P(z) and Q(z)

xc 1

odd is 2

even is 12

2/)1(2/)1(

22/

kxcc

kcc

kk

kk

odd is 2

even is 12

2/)1(2/)1(

22/

kxcc

kcc

kk

kk

12 22 xc

xxxccc 342 3123

18812 24224 xxcc

xxxxccc 520162 35235 ...

/ 2 11

/ 22 2 21

( ) cos( ) cos( )m

m mi m

i

P p i p

/ 2 1

1/ 22 2 2

1

( ) cos( ) cos( )m

m mi m

i

P p i p

Page 92: More On Linear Predictive Analysis 主講人:虞台文. Contents Linear Prediction Error Computation of the Gain Frequency Domain Interpretation of LPC Representations

Find the Roots of P(z) and Q(z)

xc 1

12 22 xc

xxxccc 342 3123

18812 24224 xxcc

xxxxccc 520162 35235

Consider m=10.

54321 cos2cos3cos4cos5cos)(~

pppppP 54321 cos2cos3cos4cos5cos)(~

pppppP

/ 2 11

/ 22 2 21

( ) cos( ) cos( )m

m mi m

i

P p i p

/ 2 1

1/ 22 2 2

1

( ) cos( ) cos( )m

m mi m

i

P p i p

Page 93: More On Linear Predictive Analysis 主講人:虞台文. Contents Linear Prediction Error Computation of the Gain Frequency Domain Interpretation of LPC Representations

Find the Roots of P(z) and Q(z)

xc 1xc 1

12 22 xc 12 2

2 xc

xxc 34 33 xxc 34 3

3

188 244 xxc 188 24

4 xxc

xxxc 52016 355 xxxc 52016 35

5

11 2 3 4 52( ) cos5 cos 4 cos3 cos 2 cosP p p p p p 1

1 2 3 4 52( ) cos5 cos 4 cos3 cos 2 cosP p p p p p

5 3 4 21

3 2 12 3 4 52

( ) (16 20 5 ) (8 8 1)

(4 3 ) (2 1)

P x x x p x x

p x x p x p x p

215 3 1 4 2 3 12

3 4 52 1

( ) ( ) ( 3 5) (2 8 )

(4 20) 8 16

P p p p p p x p p x

p x p x x

215 3 1 4 2 3 12

3 4 52 1

( ) ( ) ( 3 5) (2 8 )

(4 20) 8 16

P p p p p p x p p x

p x p x x

Page 94: More On Linear Predictive Analysis 主講人:虞台文. Contents Linear Prediction Error Computation of the Gain Frequency Domain Interpretation of LPC Representations

Find the Roots of P(z) and Q(z)

1)( 1011 aap

1922 )( paap

2833 )( paap

3744 )( paap

4655 )( paap

215 3 1 4 2 3 12

3 4 52 1

( ) ( ) ( 3 5) (2 8 )

(4 20) 8 16

P p p p p p x p p x

p x p x x

215 3 1 4 2 3 12

3 4 52 1

( ) ( ) ( 3 5) (2 8 )

(4 20) 8 16

P p p p p p x p p x

p x p x x

Page 95: More On Linear Predictive Analysis 主講人:虞台文. Contents Linear Prediction Error Computation of the Gain Frequency Domain Interpretation of LPC Representations

4655 )( qaaq

Find the Roots of P(z) and Q(z)

215 3 1 4 2 3 12

3 4 52 1

( ) ( ) ( 3 5) (2 8 )

(4 20) 8 16

Q q q q q q x q q x

q x q x x

215 3 1 4 2 3 12

3 4 52 1

( ) ( ) ( 3 5) (2 8 )

(4 20) 8 16

Q q q q q q x q q x

q x q x x

1)( 1011 aaq

1922 )( qaaq

2833 )( qaaq

3744 )( qaaq

Page 96: More On Linear Predictive Analysis 主講人:虞台文. Contents Linear Prediction Error Computation of the Gain Frequency Domain Interpretation of LPC Representations

Find the Roots of P(z) and Q(z) 21

5 3 1 4 2 3 12

3 4 52 1

( ) ( ) ( 3 5) (2 8 )

(4 20) 8 16

P p p p p p x p p x

p x p x x

215 3 1 4 2 3 12

3 4 52 1

( ) ( ) ( 3 5) (2 8 )

(4 20) 8 16

P p p p p p x p p x

p x p x x

215 3 1 4 2 3 12

3 4 52 1

( ) ( ) ( 3 5) (2 8 )

(4 20) 8 16

Q q q q q q x q q x

q x q x x

215 3 1 4 2 3 12

3 4 52 1

( ) ( ) ( 3 5) (2 8 )

(4 20) 8 16

Q q q q q q x q q x

q x q x x

cosxWe want to find i’s such that

0)(~ iP 0)(~ iP

0)(~

iQ 0)(~

iQ

Page 97: More On Linear Predictive Analysis 主講人:虞台文. Contents Linear Prediction Error Computation of the Gain Frequency Domain Interpretation of LPC Representations

Find the Roots of P(z) and Q(z)

Algorithm:

We only need to find zeros for this half.

0

Page 98: More On Linear Predictive Analysis 主講人:虞台文. Contents Linear Prediction Error Computation of the Gain Frequency Domain Interpretation of LPC Representations

0

)(~

iQ

Find the Roots of P(z) and Q(z)

2 3