ee225d spring,1999 medium & high rate coding* voice - excited channel vocoder * velp & relp...

EE 2 CTURE ON MEDIUM AND HIGH RATE CODING

N.M 26.1

PE Spring,1999

25D LE

ORGAN / B.GOLD LECTURE 26

University of CaliforniaBerkeley

College of EngineeringDepartment of Electrical Engineering

and Computer Sciences

rofessors : N.Morgan / B.GoldE225D

Medium & High Rate Coding

Lecture 26


N.M 26.2

f Other Sounds.

ms

CM)-(CVSD)

25D LE


Quality, Robustness, Speaker I.D., Possible Vocoding o

Summary for Chapter 33 - Medium & High Rate Syste

* Voice Excitation & Spectral Flattering

* Voice - Excited Channel Vocoder

* VELP & RELP

* Wave form Coding using predictive methods.(ADP

* APC

* Subband Coding

Emphasis in an getting a good efficient repesentation of the excitation.


N.M 26.3

on APC.

CELP.

on

t on Error Signal.

25D LE


*Multi pulse LPC

* CELP Celp is a variation

* Reducing Codebook Search Time in

* Back Ward Filtering

* Multi Resolution Codebook Search

* Partial Sequence Elimination

* Tree Structured Delta Codebooks

* Adaptive Codebooks

* Linear Conbination Codebooks

* Vector Sum Excited Linear Preducti

* Adaptive Transform Coding

Representation

Accen

of good Excitation

is still the basic issue.


N.M 26.4

ital System

Pure Waveform Coding

nel basis.

rom same

mited Excitation Function.

of the speech.

+

25D LE


Different Structures for Wide Band & Medium Band Dig

* Pure Waveform Coding Subband Coding

on a channel by chan* Pure Modelling

* Hybrid Systems

* Voice - Excited System Excitation derived f

* Processing of the Error Signal to Produce a Band Li

APCCELPMulti-Pulse

Many Variation

band limited function

speech

Waveform Coding

0 f1–

f1 f2– Waveform Coding


N.M 26.5

PP CUSD

ntization.

fn 1+( ) f∆

int, fewer bits are pared to I.

sing!, so original can be recovered.

25D LE


Pure Waveform Coding(Very Robust)

lay Molly’s tape.ossible Viewgraphs : Local max construct ADPCMFig.33.8, 33.9, 33.10.

Sampling & Qua

I

IIf

f f∆

n f∆

f∆

B

Let’s stick to pure PCM.from a perceptual viewponeeded to encode II, com

Each band is sampled and quantized.

Important Point : if bandwidths are judiciously clean, sampling can be done at Nyquist rate. fs 2B=

No alia


N.M 26.6

nt, fewer bits

s are judicialy clean,

ompared to I.

Nyquist rate.

inal can be recovered.

25D LE


Subband Coding

Let’s stick to pure PCM.

From a perceptual Viewpoi

Each band is sampled and quantized.

Important Point: If bandwidth

are needed to encode II, c

sampling can be done at B

I

II

f

f

f

No Aliasing! So orig

∆f

n∆f n 1+( )∆f

fs 2B=f

f

∆f

Fig.33.14


N.M 26.7

0’s)

eech Signal (0-8kHz)

Hz].

der was NOT Robust.

generally carried

annel vibrations.

less sounds.

25D LE


Voice -Excited Channel Vocoder (late 1950’s to early 196

Motivations :

(2) Possible Transmission of a Wide Band (Analog) Sp

throughan excisting Telephone Channel [300-3000

(1) General Faling - Excitation Signal in Channel Voco

* Intuitive grasp of the fact that baseband [0-900Hz],

all the necessary excitation information for vocal ch

* A white noise same use a suitable excitatin for voice


N.M 26.8

Base Band

er Synthetic sizer Speech

Speech

tation spectrum.

nerally carried

nnel vibrations.

ess sounds.

25D LE


Original Idea

Analysis

Synthesis

Problem :

Speech Channel Spectral

100-900Hz

100-900Hz Non-Linear Vocod

Base Band

Vocoder Analyzer

Filter

Distoration Synthe

Speech Signal

Base Band Speech

100-800Hz

Distortion of the exci

* Intuitive grasp of the fact that baseband [0-900Hz], ge

all the necessary excitation information for vocal cha

* A white noise same use a suitable excitatin for voicel


N.M 26.9

ced spectral flattering

ortion products of the

25D LE


In early 1960’s, Schroeder, David et al at BTL, introdu

-got rid of spectral distortion [more or less].

Base Band Signal

Simple non-linearity (rectifier)

Result of non-linearnity distfrequencies are harmonics fundamental frequency.

f

Spectral Flattening


N.M 26.10

e Predictive Coding.

d

time

25D LE


In the late 1960’s, schroeder and Atal introduced APC -Adaptiv

Long Term Predictor Prediction

Sample to be predicte

Previous Samples


N.M 26.11

n

the sampling rate.

M)

e1 n k–( ) e2+

25D LE


e1 n( )

Two types of prediction.

Perform LPC Analysis on

By transmitting

Major Assumption - is so small

that it can be quantized to 1 bit. BUT SENT at

Predicted Valuey n( ) αy n T–( ) y n( ) e n( )+= =

so e1 n( ) y n( )– αy n –(+=

a1 a2 …ak and α M,, ,

e1 n( ) a1e1 n 1–( ) a2e1 n 2–( ) …ak+ +=

e1 n( )

e2 n( )


N.M 26.12

.

signal costs 8kbs.

2kbs.

Bit Rate.

is big.

25D LE


Even a ONE bit error signal results in a large bit rate

If sampling rate is 8kHz, then transmission of error

Addition of transmitting ’s sould be another

Early APC Systems Operated at 9600bps.

Major Research Efforts : Reduction of Error Signal

If pitch is wrong, first error signal

d M ak, ,

e1 n( )Pitch Detection


N.M 26.13

dard Excitation Signals.

mpling rate.antization.

pefully].

25D LE


* LPC - Error Signal is eliminated and replaced by stan

* RELP - Residual Excited linear Prediction.

(like Channel Vocoder)

Low Pass Filter of error Signal - reduced sastill one bit qu

* VELP - Voice-Excited LP.

So error signal rate can be reduced [ho

They aimed for 4800 bps.

Key to Error Signal Reduction.


N.M 26.14

vailable for

the Error Signal.

vailable.

ter.

sequentially

n frames] and

tain a GOOD fit

H.countered this ideaSetevens Halle concept pter 17.

25D LE


As computer speeds increased, New sptions became a

real-time Coding of

Basic Philosophy - Analysis by Synthesis

* Transmitting system has both analyzer and synthesizer a

* So Synthetic Speech can be generated at the transmit

* Using same criteria, the synthetic speech is compared

with the actual speech [perhaps every frame, or every

synthesizer parameters obtained by analysis VARIED to ob

between ACTUAL SPEECH vs. SYNTHESIZED SPEEC

* So the best parameters are sent. we enin the in cha


N.M 26.15

r quantized error signal.

A

Synthetic

act Parameters

uered difference

esisSpeech

s n( )

he Vocal Tract Parameters.

R meters

Synthetic SpeechTe

��

25D LE


Basic Idea - Replace the one-bit error signal of APC with a vecto

nalyzer

Store of

Speech LPC

LPC

Vocal Tr

mean sq

M Sequences, each of length

one frame.

Change

Address

Analyzer

Synth

s n( )

Transmit address of best of the M sequences transmit t

eceiver

LPC Synthesis

Store of M Sequences, each of length

one frame.

Vocal Tract Para

“best” exciter

best address

his is an

…

xample of VQ.

��


N.M 26.16

lyzer must

t speaker]

es in a single frame.]

stems can opearate

25D LE


Problems discussed in Chapter 33

* How to create the stone?

* Perceptual weighting filter.

* Delay

* Reduction of Codebook Search. [In 20ms, the ana

* Adaptive Coding.[ Storage is adaptive to a differen

performe complete analysis-synthesis many tim

So computer speed must be such that MANY sy

in real time SIMULTANEOUSLY.


N.M 26.17

s(z 1–

s n( )

Modulation.

s( n( )

z 1–

s n( )+_

V

25D LE


_

n) e n( ) e n( )

e n( )z 1–

Figure 33.7 : Differential Pulse Code Modulation.

Figure 33.8 : Adaptive Differential Pulse Code

Quantized error Signal

Adaptive Format

Low-Pass Rectification

+Q

_n) e n( )

e n( )e

z 1–

Filtering

Quantizer Q(V)

Format

V


N.M 26.18

s n( )

Receiver

M Analyzer

Σ

z 1–

z 1–

1

M n( )

a1 0.9418=

odulation.

25D LE


Channel

M

s n( ) e n( ) e n( )

s n( ) a1s n 1–( ) M n( )+=

Analyzer

Transmitter

Σ Hard Limiter

z 1–

z 1–

Σ

z 1–z–

…

s n 1–( )

M n( ) f(e n( ) e n 1–( ),,=

e n 2–( ))a1

a1

Figure 33.9 : Continuously Variable Slope Delta M


N.M 26.19

S n( )

oncept

S

z( )

11 P z( )–--------------------

25D LE


s n( )

P z( ) e n( )

+Q

Figure 33.10 : Rudimentary Linear Prediction C

peech

AnalyzerParameter

+

Coder

P

Quantized Error Signal1 P z( )– +

_


N.M 26.20

drature Mirror Filters.

s' n( )

K1

_

K21:2

1:2

25D LE


Figure 33.14 : Four Channel Subband Coder with Qua

_

e1 n( )

x n( )

+

H2

K1

K2

e2 n( )

e4 n( )

H1

+

H1

_

+ K1

e3 n( )

H1

+

K2

H2

H2

1:22:1

2:1

2:1

2:1

2:1

2:1

1:2

1:2

1:2


N.M 26.21

25D LE


Figure 33.2 : Zig-Zag Network

OUTPUT


N.M 26.22

ignal

Frequency

25D LE


Figure 33.1 : Spectral Flattening of the Base-Band S


N.M 26.23

ning.

tic

S

Fig. 33.1c

25D LE


Figure 33.3 : Implementation of Spectral Flatte

Hand Limiter Input-Output Characteris

peechLow-Pass

Half-Wave

Bandpass Hand

Fig. 33.1a

Fig. 33.1b

Output

Input

Filter

Rectifier

Filter

Limiter

Bandpass Hand Filter Limiter

Bandpass Hand Filter Limiter


N.M 26.24

Coding

to RCVR

)

and

Packing

25D LE


Adaptive Predictive Coding of Speech

Input One Bit

LPC

Pitch

First Error Signal

Σ

-

+

αzM–

sn

a1

e n(

α

akzk–∑

M

q

Quantizer

Σ

Σ

Σ

Σ

αz

M–Analysis

Analysis

a2

a4

q

Speech


N.M 26.25

I

Pass

ignal

Spectrum

Output r

Flattener

+S Speech

tion)

Signal

Spectrum

Output

Flattener

S Speech

25D LE


Block Diagrams of VELP and RELP.

nput

LPC

Low-Pass

Coding LPC High-

Excitation S

Distotion

Analysis

Filter

Synthesis

Network

Filte

Low-Pass Filter

Formatting

Transmission

Decoding

peech

(a) VELP (Voice-Excited Linear Prediction)

(b) RELP (Residual-Excited Linear Predic

LPC parameterInput LPC

Low-Pass

Coding LPC

Excitation

Distotion

Analysis

Filter

Synthesis

NetworkLow-Pass Filter

Formatting

Transmission

Decoding

peech Error Signal


N.M 26.26

hm.

25D LE


Spectra at different nodes in the VELP algorit


N.M 26.27

g Delay Predictors.

I

Speech

Selecting

Instantaneous_ Objective

Error

igianl eech

gnalSn

25D LE


Figure 1 : Speech Synthesis Model with Short and Lon

++nnovation

Fine Structure Spectral Envelope

Long Delay Short Delay

(Formants) (Pitch) “White”

Predictor Predictor

Figure 2 : Block Diagram Illustrating the Procedure for

++Innovation

Long Delay Short Delay

“White”

Predictor Predictor

PerceptualAverage Square

Perceptual Weighting Filter

Error

Or Sp Si

Sn

the Optimum innovation Sequence.


N.M 26.28

25D LE


Analysis by Synthesis

How to create the store?

Reduction of delay.

Reduction of Codebook Search

Adaptive Codebook.

APC 10,000bps.

Bessie Smith

Louis Armstong.


N.M 26.29

e1 n( )

25D LE


APC - Adaptive Preductor Coder Late 1960’s

pitch

Analysis

LPC

M

e1 n( )

e2 n( )

y n( ) αy n M–( ) y n( ) e1 n( )+= =

e1 n( ) y n( )– αy n M–( )+=

e1 n( ) a1e1 n 1–( ) a2e1 n 2–( ) … e2 n( )+ + +=

i bit

y n M–( ) y n( )


N.M 26.30

ocodar

to get Flat Spectrum.

nthesizer

Hz

ts per second.

25D LE


Channel

Low-Pass Baseband

V

Non-Linear

Impossible

Distortion

Sy

Filter 0-800Hz

Vocoder Analyzer 800-8000Hz

Parameter

0-800

10,000 bi


N.M 26.31

1 f2

f∆ n 1+( ) f∆

25D LE


Subband Coding

64k bits

20-30k bits

4bits

2bits

0 f1–

f1 f2–

fn fn 1+–

0 f1–

f1 f2–

1500

4000

f

n


N.M 26.32

r than Speech

25D LE


Medium & High Rate Systems

Quality

Robustness Speaker I.D.

Possible Vocoding of Sounds othe

Waveform - PCM - 8kHz, 8bits 64kbits.(Subband Coding)

DPCM ADPCM CVSD

Basic Issue

Efficient Representation of the Excitation.

Voice-Excited Systems LPC

Channel Vocoders

Prediction - APCCELPMulti-Pulse

ee225d spring,1999 medium & high rate coding* voice - excited channel vocoder * velp & relp...

Documents