EE225D Spring,1999 Pitch Detection &Vocoders · but we saw in Chapter 16 that pitch would be...

EE 2 TURE ON PITCH DETECTION AND VOCODERS

N.M 23.1

PE Spring,1999

25D LEC

ORGAN / B.GOLD LECTURE 24

University of CaliforniaBerkeley

College of EngineeringDepartment of Electrical Engineering

and Computer Sciences

rofessors : N.Morgan / B.GoldE225D

Pitch Detection &Vocoders

Lecture 24


N.M 23.2

)

resentation?

e] and

25D LEC


Major Question

How to make a “Perfect” Vocoder? (Can it be done?

What limitations are encountered for low bit rate rep

Today’s Topic

Traditional 2400bps systems [or at least in that rang

Pitch &Voicing detection.

NEXT

Very low rate systems [600bps]

Higher quality more rubust systems at 5-30Kbps

NEXT


N.M 23.3

resu lt

the sam e stim u lus.

n ta l frequency estim ato r”

ived” even if

ve descrip tion

25D LEC


D ifficu lties E ncountered in P itch D etection

* Purpose o f p itch detection is to au tom atica lly obta in a

that is in agreem ent w ith a psychoacoustic resu lt for

* E arly researchers p referred to use the term “ fundam e

but w e saw in C hap ter 16 that p itch w ou ld be “perce

the stim u lus w as a harm on ic for that frequency.

(exam p le - sh ift o f v irtua l p itch )

* W hat w e’re really after is the N AT U R E and quantita ti

A nd a lso to m ake a vocoder sound natu ra l.

o f the exc ita tion function.


N.M 23.4

ive.

bove.

nalysis d iff icu lt.

s

m e adu lts ch ildren 16 :1 range

ting track ing the period .

l tract constric tion .

25D LEC


3 . R epresen tation o f the transien t exc ita tion du ring p los

4 . R epresen tation o f the no ise for a w h ispered vow el

5 . R epresen tations o f various com b itnations o f a ll the a

E xpam p les o f speech w aveform s that m akes the above a

* D ynam ic range of quasi-period ic vocal co rd v ib ration

as low as 50H z fo r soas h igh as 800H z fo r

* R ap id varia tion in g lo tta l period* S udden change in vocal treat shape [e.g . nasal]* Transition from unvoiced to vo iced .* E nv ironm enta l transm ission p rob lem s.

* T h is m eans: 1 . D etection o f the tim e w hen the vocal cords are v ibra

in a [perhaps rap id ly vary ing ] quasi-period ic w ay and

2 . R epresen tation o f the fr ic tion no ise caused by a voca


N.M 23.5

W ted as the

lter function .

* and E ω( )

ned as fo llow s.

*may be different.

* are time-varying.

25D LEC


ith linear assum ptions, the speech w ave can be rep resen

C onvo lu tion of an exc ita tion function w ith a vocal tract f i

In spectra l term s : S ω( ) E ω( )H ω( )=

In a channel vocoder analyzer, m easurem ents , H ω( )S ω( )

are N O T com puted separate ly.

* In the channel vocoder syn thesizer, the spectrum is ob ta i

If E ω( ) is a FLAT SPECTRUM , S ω( ) S ω( ),≅ although the phases

The situation is complicated by the fact that E ω( ) and H ω( )

W hite N o ise

Pu lse G enerato rR outer

e t( )

S ω( ) E ω( )H ω( )=

G enerator


N.M 23.6

tion s ignal.

n

th

sate by being

)

25D LEC


In L PC , w e start o ff w ith

change o f nom enclatu re is the m odel o f the speech excita

L PC derives an all-pole m odel

It would be nice if Hˆ ω( ) was really a good representatio

Speech can be perfectly reconstructed by carvolving h n( ) wi

the error signal e n( ).

correspondingly different than Ex ω( ) .

if H ω( ) differs greatly from H ω( ), E ω( ) will compen

s n( ) e n( ) h n( )×=

S ω( ) E ω( ) H ω( )⋅ Ex ω( ) H ω(⋅= =

S n( ) Ex ω( ) H ω( )⋅=s ω( ) ex n( ) h n( )⋅=

ex n( )

of H ω( ) , the real vocal tract function.

H ω( ) h n( )→


N.M 23.7

f ilter separation

tra l dom ain ,

y, the m odel assum es

on , the exc ita tion

nted and then

ow er

eparation .

25D LEC


H om om orph ic analys is has the hypo thesis that source -

T he m odel a lso assum es that these are m u ltip lied in the spec

so that tak ing the log turns the product in to a sum . F inall

that the tw o are separab le w ith lif tering. G iven th is separati

function and the vocal tract f ilter function can be represe

C onvo lved to g ive the syn thesized speech.

H ω( )

* M any L P C system s [m u lti pu lse, ce lp , e tc .] derive their p

by search ing for an erro r s ignal that com pensates for .

is m anifested as spectrum envelope - spectra l f ine structu re s


N.M 23.8

), a ll system s re lay on the

b le period pu lse source

c tra and take few b its

0bpss

25D LEC


In o rder to ach ieve low transm ission rates (e.g . 2400bps

excita tion m odel consisting o f a no ise source and a varia

- B oth sources are reasonab le approx im ations to f la t spe

buzz-h iss sw itch - 1b it every 10m sec. 10pitch tracker - 6b its every 10m sec. 600bp

to transm it.


N.M 23.9

M resent, F u ture .

cam e a standard.

e quality at 2400 bps.

o liferate,

nger the so le criterian .

tradeo ff.

m ethodsPC M ,A D PC Mictive ab ility

2

25D LEC


a jor M otiva tion for D orry R esearch on Vocoders : P ast, P

P ast - S ecrecy - W W II - D ata rates w ere lim ited . 2400bps be

N early a ll fund ing cam e from D O D to try to im prov

P resen t - M odem s are m uch better. A s cellu la r phones p r

date rate l im itations sti ll app ly but 2400bps is no lo

M ain d irection is sti ll quality (robustness) - b it ra te

F uture - G reater robustness - e ffic ient sto rage o f

speech (and m usic) - coding -recognition tie-in .

Tw o S ides of the C o in

B asic M odels fo r A nalysis & S yn thesis

channelLP CH om om orph ic

Wave fo rmPC M ,A

S om e p red

1


N.M 23.10

ction and

System

l Tract

menting a

.

s.

25D LEC


Complete Channel Vocoder

* Synthetic speech is the convolution of an excitation fun

a vocal tract filter function.

* Assumption : Synthesizer is a Time variable Linear

If this assumption was wrong and excitation and Voca

Interacted in some Non-Linear Way, problem of imple

“transparent” system probably becames intractable

Working Hypothesis for 2400bps (and lower) system

Excitation is either buss [variable pulse generator]

Remember basic assumption for all vocoders.

or hiss [white noise generator]


N.M 23.11

hat tracks perfectly and

ator at the synthesizer is

ents.

pectral distortionintroduced bypitch jitter.

envelope caused by jitter

ω

25D LEC


Now, assume that you have built a great pitch detector t

records

Now, this information is transmitted and the buzz gener

forced to produce pulses based on the above measurem

Most of the time,

T1 T2 T3 etc., , ,

Spitch dose NOT behave this badly.

Consider the spectrum of a jittered pulse train.

E ω( )T4T1 T3T2

time


N.M 23.12

20msec.. Analyzer

a period of 20msec.

not as bad as

bove E ω( ) and H ω( ).

ω) E ω( ) H ω( )⋅=

Spectral distortionintroduced bypitch jitter.

25D LEC


In real life, lets assume that analysis takes place every

generates a single pitch number, so at synthesizer, for

actual excitation during voicing. ~

20ms. 20msec.

time

time

T1 T2 T3 T4

S ω( ) is the product of the a

S(

e t( ) s t( ) S ω( ) E2 ω( ) H ω( )⋅=

at the synthesizer


N.M 23.13

oise.

e

ls

Synthetic speech

) H ω( )⋅

with Spectrally flattened

excitation.

25D LEC


Spectral Flattering

Turn the excitation signal into a white signal or white n

xcitation

Modulation signa

Model of S ω( ) E ω(=

+

B.P.Filter

Hard Limitor

B.P.Filter


N.M 23.14

elope function

ctrally flattered.

rmine perceived quality.

spectral flattering

sizer.

ants”.

nvelope function?

25D LEC


Major Question

Does all-pole synthesizer model the Vocal tract env

- if the former is true, excitation should NOT be spe

- if the latter is true, spectral flattering may help.

* Joe Tierrey and I did an informal experiment to dete

The result was ambiguous.

* In general, existing LPC systems (low rate) do NOT use

It may depend on the ORDER of the pedictor & synthe

a 10th order predictor correspends to five “form

or the complete speech e


N.M 23.15

el vocoders &LPC.

hink] been tried but

PC).

pe could be completely

e

lope

ght to be perfect.

25D LEC


Homomorphic Vocoder

* Excitation is modelled in the same way as for chann

Spectral flattering of the exciation signal has never [I t

it should work ( in the same ballpark as channel &L

Point C is Cepstrum.

Vocal tract

Low time High time cepstrumNote - if excitation and envelo

tim

High time comjienent of enve

Envelope

Speech Spectrum Envelope

or

cepstrumseparated, synthesis ou


N.M 23.16

ion Spectral

Pattern

e Function

Recognition for Pitch

25D LEC


512-Point Log Separat

Excitation

FFT Magnitude in Tim512-Point

FFT

(a) (b) (c) (d)

Figure 20.1 : Cepstral analysis.


N.M 23.17

orrelation Function

ed Speech.

ections with 15ms

25D LEC


TIM

E

L A G 15m s

Figure 30.8 : Autoc

of Spectrally Flatten

Successive 30ms s

overlap.


N.M 23.18

O u tpu tΣ

n An⁄

1 A1⁄

peech S ignal.

F la ttened

O rig inal

25D LEC


Input

B PF

FW R Sm ooth

D elay ÷F1

Fn

S1

Sn

A1

An

Cn S=

C1 S=

B PF

FW R Sm ooth

D elay ÷

F igure 30.7 : Spectra l F la tten ing and its E ffect on the S


N.M 23.19

ntial Decayime)

25D LEC


Figure 30.3 Extraction of the Period

Variable Blanking Time Variable Expone(Rundown T

Firings


N.M 23.20

25D LEC


Figure 30.6 : Low-Pass filtered speech signal.


N.M 23.21

h Detection.

ge of Pitch

ns in Time

ariations in Time

iced Transition

peech

ise Background

25D LEC


Figure 30.4 : Six Examples of Difficulties in Pitc

Dynamic Ran

Pitch Variatio

Vocal Tract V

Voiced-Unvo

Telephone S

Acoustic No


N.M 23.22

D etection.

25D LEC


F igu re 30 .10 : C epstra l A nalys is for P itch

40dB

time

time


N.M 23.23

el.

h

S

Pitch Detection

Stage 2

Based on Correlation

with Reference Patterns

25D LEC


Figure 16.9 : Block Diagram of the Periodicity Mod

Figure 16.10 : Block Diagram of the Place Model.

SignalG lobal

Pitc

Filter Bank Elementary

ignal 4096-PT Detection

Pitch Detection

Stage 1

F1

F2

FM

EPD1

EPD2

EPDM

Pitch Detectors

P itch D etec tion A lgo r ithm

FFT Spectral Peaks

Based on Separation

Between Peaks

of


N.M 23.24

inner

Freuency

p5 p6 Quanta

ude

25D LEC


Figure 30.13 :

Frequency

1050Hz

W

p1

p1

p1 p1p2

p2

p2

p2

p2

p3

p3

p3 p3

p3

p3

p3 p3

p4

p5

p5

p6

p1 p2 p3 p4

p n( )

Spectral Magnit

12

345 67

Armonic Pitch Detection Algorithm


N.M 23.25

ndidate (P.C.) 1

P.C. 2frequency (Hz)

P.C. 3

P.C. n

lgorithm

ctrum

25D LEC


Pitch Ca

Figure 30.14 : Goldstein-Duifhuis Optimum Processor A

Spe


N.M 23.26

sizer

Magnitude

Signals

Signals

Encode

Encode

Signals

TRANSMIT

25D LEC


Figure 31.2 : Channel Vocoder Analyzer and Synthe

V/UV

PitchPitch

Voicing

Detector

Detector

Bandpass Magnitude

Lowpass DecimateFilter N Filter N

x n( )

Bandpass Magnitude

Lowpass DecimateFilter 1 Filter 1


N.M 23.27

sizer

Vocoder

Output+

25D LEC


Figure 31.2 : Channel Vocoder Analyzer and Synthe

Switch

Pulse Generator Noise Generator

Bandpass Filter 1

Bandpass Filter 2

Bandpass Filter N

RECEIVE

V/UV Signals

Pitch Signals

Magnitude Signals


N.M 23.28

Estimate

Half-Wave Rectifier.

25D LEC


Figure 31.4 : Effect of Pitch Ripple in a Spectral

Figure 31.3 : Example of Energy Measurement With a


N.M 23.29

r to Modem

Synthesis

C

Synthesized

M

esizer

Parameters

Speech

thm

25D LEC


Input Measurement of Computation of Paramete

Pitch and Voicing

from Parameter

Buss Generator

Hiss Generator

LPPitch Voicing

Gain

Autocorrelation Function

Symthesis Parameters Coding

Detection

Decodingodem

SynthBit

Figure 31.11 : Block Diagram of the LPC Algori


N.M 23.30

age 0

In

B Front (Lips)

Output

z 1–

1

K0

K0

-

25D LEC


Figure 31.12 : Lattice Synthesizer for LPC

+

Stage P-1 Stage 1 St

put

ack (Glottal)

... +

z 1–

+

+ + +z 1– z 1–

-

+

-+

+ +

++...

...

Kp 1– K1

Kp 1– K1

- -

-


N.M 23.31

verse

nvolveSynthetic

ourier nsform

Speech

25D LEC


Schematic of the Homomorphic Vocoder.

Input Speech Samples

Fourier

Log

InMultiplier

Cepstral

Fourier Exponential

Co

Excitation

Excitation Parameters

Analyzer Synthesizer

Transform

Magnitude

FTra

Generator

Pitch Detector

Transform

Inverse Fourier

Transform

Time

1

C


N.M 23.32

is

ultiplier

ime

c nT( )

h nT( )

c nT( )

s nT( )

e

e

25D LEC


Figure 31.13 : Cepstral Vocoder Analys

Log Inverse MFourier

MagnitudeFourier

TransformTransform

T

s nT( ) S ω( ) s nT( )

S ω( )

s nT( )

Time

FrequencyTim

Tim


N.M 23.33

scan in

25D LEC


Inverse Fourier Exponential

Convolve

Excitation Excitation Parameters

Fourier Transform

Generator

Transform

Figure 31.14 : Cepstral Vocoder Synthesizer

c nT( )v nT( )

e nT( )

EE225D Spring,1999 Pitch Detection &Vocoders · but we saw in Chapter 16 that pitch would be...

Documents

Transcript of EE225D Spring,1999 Pitch Detection &Vocoders · but we saw in Chapter 16 that pitch would be...