EE225D Spring,1999 Low Rate Coding · ee 225d lecture on low ra te coding N.MORGAN / B.GOLD LECTURE 23 25.3

EE 2 LECTURE ON LOW RA TE CODING

N.M 25.1

PE Spring,1999

25D

ORGAN / B.GOLD LECTURE 23

University of CaliforniaBerkeley

College of EngineeringDepartment of Electrical Engineering

and Computer Sciences

rofessors : N.Morgan / B.GoldE225D

Low Rate Coding

Lecture 25


N.M 25.2

Algorithms

3 fferential

1

Modelation [Standard?]

ystemslling,rm Coding)

d Channel Vocoders LPedictive Coding.

ulars, STCders, LPC, Cepstral.

zation Algorithms.

Standard 2400bps Systems.n to Reduce arameters

tization

-Synthesiscoder.

25D


Digital Vocoder Date Rates Applications & Performance

0,000 bps

5,000-10,000 bps

2,000-5,000 bps

Wide-band, High Fidelity

Medium Band, Good Quality

Reasonable Quality

More Vulnerable to Environment.

Adoptive Di

1,000-2,000 bps

500-1000 bps

100-500 bps

2,400 bps (standard)

0,000-20,000 bps

Speech Transmission

Speech Transmission.Some Noise Immunity

Speeker I.D.? Secrecy.Pitch Errors Restricted Bandwidth.

Secrecy.

Extremely Restricted

Underwater Transmission?

Pulse Code

Split Band S(Part ModePart Wavefo

Voice ExciteVELP & REAdaptive Pr

Telephone Speech

CELP, MultipChannel VocoSpeech Diziti

Frame-fill for Transformatio# of Bits for P

or Busy Channels.

Super-Restricted Channels.

Vector Quan

RecognitionPhonetic Vo

Very Restricted Channels.


N.M 25.3

y be assigned

ore bits.]

25D


Frame - fill

* Start with a 2400 bps algorithm [Channel Vocoder].

* How do you get 2400bps?

Assume 400bps for excitation.

Assume 20ms frame or 50 frames/sec.

20005∅

------------ 40 bits/channel for a channel vocoder.=

Assume 16 channels 4016------ 2.5 bits/channel, bits ma=

on a perceptual basis. [Usually low channels need m


N.M 25.4

.

be reduced.

t B A.=

t B C.=

onstruct

C.

1050

25D


Question

How to reduce bit rate below 2400bps?

Algorithm- Transmit A&C

- Reconstruct B at Receiver

Control bits tell Receiver

So 40bits gets reduced to 22. Pitch can also

Measurement are model

- If B A reconstruc,≈

- If B C reconstruc,≈

-If B A≈ B C, rec≈∉

A B C

Channel nn

B by averaging A ∉

40bits 50× 2000= 42bits 25× =

at the transmitter.

how to handle the incomming data.


N.M 25.5

e same structure

re a few different

st structure for

sult

works best.

25D


Frame - fill for LPC

Whereas the channel vocoder synthesizer is always th

(a filler bank with malulaters for each b.p.filter) There a

LPC structures. For frame-fill, the trick is to find the be

finding a criterion of similarity.

a-parameters direct formk - parameters lattice formarea rationlog area ration

Experimental Re

Lattice structure

DRT Results

Table IV?


N.M 25.6

able.

d transmitter.

ns.

by 2:1.

ble, we

25D


Pattern Matching on Vector Quantization

Consider the 2400bps channel vocoder.

Assume 40bits per frame.

* This corresponds to 240 possible patterns.

Assume that 106 patterns are perceptually distinguish

If all 106 patterns were stored at both the receiver an

Algorithm

- Compare incoming pattern with All 106 stared patter

- Send the 20 bit address of the best stared pattern.

- Reconstruct this best pattern at receiver.Thus, bit rate is decreasing

If only 1000 patterns were perceptually distinguisha

gain from 40 to 10 bits per frame. [4:1 Reduction]

C.P. Smith (1960’s) Buzo et al (1980’s ?)


N.M 25.7

rs.

lter.

igrates towards unit circle.

d.

o transmit.

icedrmantsRCOR

25D


Kany - Caulter 600bps Vocoder

Includes LPC analysis

clever LPC parameters to fewer formant paramete

Vector quantization. - a form of frame fill.

Viewgraphs

Fig. 32.1 - Complete block diagram of Kamy - Cou

Fig. 32.2 - loci of poles is

Fig. 32.3 - Spectrum of all pole system as poles m

Vector Quantization - 128 formant patterns were store

Figure overall rate,

possible parameters t(7bits)

(including everything) to get 600 bps.

K10 1→

VofoPA

Viewgraph of Lattice


N.M 25.8

Control bits are sent to

frames. The sent frames

nd decides whether these

e than just two adjacent

ion rate is Variable, and

r quantized.

25D


Frame - Fill vs. Merging

In frame - fill, alternate frames are not sent - instead,

direct the receiver to “optimally” recreate the missing

can still be vector quantized.

In merging, the analyzer compares adjacent frames a

frames can be merged into a single frame. Perhaps mor

frames can be merged. This means that the transmiss

Control bits are needed. Merged framed can also be vecto


N.M 25.9

Synthesized

Speech

25D


Channel Signal 1

Channel Signal 2

VARIABLEBANDPASSFILTER 1

Channel Signal k

VARIABLEBANDPASSFILTER 2

VARIABLEBANDPASSFILTER k

Formant 1

Formant 2

Formant k

Figure 29.8: Parallel formant synthesizer.

Excitation


N.M 25.10

+

25D


Pu lse

N o ise Para lle l S tructu re for

M ost C onsonants

Serial S tructure fo r Vow els and N asalsSource

Source+

+

× ×

× ×

A1

A2

A3

A4

F igure 29.12 : S tructure o f K la tt Synthesizer.


N.M 25.11

Fant. Form [20]

G

etwork

NG

Σ

orkF1F2 F5

N1

25D


Figure 29.2: OVE II Speech Synthesizer of Gunnar

Pulseenerator

Fricative and Plosive N

SourceFilter 1

oiseenerator

Vowel Netw

Nasal Network

FO

FH F3 F4

NO N3 N2 N4

KO K1 K2

SourceFilter 2

SourceFilter 3

AH

AO

AN

AC

+-


N.M 25.12

A

.

rs

ultiplexer

Transmission C and Encoder

Media

25D


nalog to

Figure 32.1 : Kang-Coulter 600 bps Voice Digitizer

Filter

Input

Linear Prediction Excitation

Excitation Paramete

Pitch Period

Amplitude

Voice/Unvoice

M

Vocal-Tract-Filter Conversion From Speech

0.1 to 3.6kHz

Digital onverter

Predictive Analysis Filter

AnalysisResidual

Linear Predictive Coefficients to

Formant Frequencies

Information

Parameters

(a) Transmitter


N.M 25.13

r

Digital to

Filter

Speech

inear

Out

0.1 to 3.6kHz

Analog Converter

edictive nthesis ilter

TrM

25D


Figure 32.1 : Kang-Coulter 600bps Voice Digitize

(b) Receiver

P u lse-Tra in and L

Excitation Parameters

Pitch Period

Amplitude

Voice/Unvoice

Demultiplexer

Vocal-Tract-Filter Conversion From

PrSyF

and Decoder

Formant Frequencies to Linear Predictive

Coefficients

Information

Parameters

R an dom -N o ise E x c ita tio n

S igna l G enerato rs

ansmissionedia


N.M 25.14

Unity.

= 0Hz

25D


Figure 32.2 : Loci of the Poles as Approachesk10

f = 2000HzUnit Circle

Origianl Pole Positions

1


N.M 25.15

��

wards Unit.Π Z Zi–( )

L 1=

n

∑=

te to unit circle.

25D


Figure 32.3 : Spectra of All-Pole System as Poles Migrate ToZn α1– Zn 1– α2Z

n 2–– … αn–

’as dn 1 Zi s migra,→


N.M 25.16

Predictor Coefficients

ts

Formants

Ratios

I OutputC

F ig ed on L PC analys is.(cont.)

25D


+x n( ) y n( )

a1

z 1–

a2

a3

ap

z 1– z 1–z 1–

y n( ) pk---kkkkkkk=

...

... ap 1–

(a) Direct-Form Digital Filter with Variable “a” Coefficien

Area

(b) Acoustic Tube with Variable Area Functions

nput

≈

≈

– AAp 1–A1

u re 29 .5 : Tw o configurations fo r a ll po le syn thesizers bas


N.M 25.17

eters

F ased on LPC analysis.

age 0

In

B Front (Lips)

Output

z 1–

1

K0

K0

-

r Coefficients

oiced Sounds.

ts.

25D


(c) All-Pole lattice Network with Variable “k” Param

igure 29.5 : Two configurations for all pole synthesizers b

+

Stage P-1 Stage 1 St

put

ack (Glettis)

... +

z 1–

+

+ + +z 1– z 1–

-

+

-+

+ +

++...

...

Kp 1– K1

Kp 1– K1

- -

-

ParcoPredictor Coefficients

Synthesis Strategy for Voiced Sounds For unv

From (a)

Vector quantization of the PARCOR Coefficien


N.M 25.18

e Digitizer

00-Bit-Per-Second

40Hz

7 bits/frame

1 bit/frame

4 bits/frame

5 bits/double frame

1 bit/double frame

30 bits/double frame

Voice Digitizer

25D


Figure 32.4 : Parameter Coding for 600bps Voic

Parameter

Coding

Frame rate

Vocal-tract-filter parameters

Excitation parameters

Voice/unvoice decision

Amplitude

Pitch

Synthronization

Total number of bits

Typical 2400-Bit-Per-Second 6

44.444Hz

40 bits/frame

1 bit/frame

6 bits/frame

6 bits/frame

1 bit/frame

54 bits/frame

Linear Predictive Encoder


N.M 25.19

tored patterns.

he actual pattern.

25D


V.Q.Algonithm

* Storage of the patterns.

Compare incoming pattern with All previously s

2N patterns stored.

* Receiver also has stored availavle.

Send the address of the stored pattern nearest to t

15


N.M 25.20

ract Parameter

Text to Speech

25D


Kang-Coulter 600bps Vocoder

Started with 2400bps

Reduce # of parameter to 4 or 5

Vector quantization

Frame-fill.

10 Vocal TLPC

has no speaker I.D.

Speech Recognizer Speech to Text

75bps


N.M 25.21

Algorithms

3

2

1 ystemsling,rm Coding)

d Vocoders

C hommerphic

lars, STCders, LPC, Cepstral.zation Algorithms.

n other Systems.

tization

-Synthesiscoder.

25D


Low Rate VocoderDate Rates Applications & Performance

0,000 bps

5,000-10,000 bps

2,000-5,000 bps

High Fidelity

Medium quality

Reasonable Quality

More Vulnerable to Environment.

ADPCM

1,000-2,000 bps

500-1000 bps

100-500 bps

,400 bps (standard)

0,000-20,000 bps

(Telephony)

Good NoiseSome Noise Immunity

Speecher I.D.? Secrecy.Pitch Errors.

Secrecy.

Extremely Restrected

Underwater

DPCM

Split Band S(Part ModelPart Wavefo

Voice Excite

Channel LP[High Intelligibits]

CELP, MultipoChannel VocoSpeech Digiti

Frame-FillTransformatio

or Busy Channels.Vector Quan

RecognitionPhanetic Vo

Very Restricted Channels.


N.M 25.22

0bps Excitationocal Tract

, - Interpolate.

A, B A.=

C, B C.=

A CB̂

F1 F2F1 F2+2

-----------------

n

25D


1000-2000bps Frame-fillstart with a complete System 240

400bps 2000bps V

Analyzer

2 Control Bits

2400bps 1200bps

Synthesizer

If neither is true

Is B A ?≈

Or is B C ≈ If B is close to

If B is close to

Or is neither true.

A

B

C

0 20ms 40ms

not sent

n


N.M 25.23

el Vocoder]

2000

25D


500-10Mbps

C.P. Smith Pattern Matching

2400bps Vocoder

2000bps Spectrum

50Hz Rate 40 Bits Per Frame

240 Different Spectrum [Chann

220 1 million

Vector Quantization LPC

50 40× =

Reduction of # of Parameters


N.M 25.24

25D


Frame-fill for LPC ?

Predictor coefficients

SensitivityΚ-Parameter [PARCOR]

Area Retios.

Log Area Ratios.

’a s ↔

Documents

EE225D Spring,1999 Low Rate Coding · ee 225d lecture on low ra te coding N.MORGAN / B.GOLD LECTURE 23 25.3