Dr. Dóra Maros Institute of Telecommunication Óbuda University Telecommunications Techniques II. Source Coding (speech) Error detection and Error correction.

Dr. Dóra MarosInstitute of Telecommunication

Óbuda University

Telecommunications Techniques II.Source Coding (speech)

Error detection and Error correction

Transmission from source to destination

Source (analog)

Destination (analog/digital)

Channel (digital)

Channel coding

Channel decoding

Sourcecoding

Sourcedecoding

-Microphone-Camera-Pressure, temperature etc sensors

-Speaker (ears)-Displays (eyes)-Alarms (see and hear!)

Bandwidth allocation

Source information

Source information Filter

Filter

A/D

A/D

Compression

Compression

Channel coding

Channel coding

Modulation Channel

How we „generate” voice?

Vocal tract (resonator filter system)

Excitation

Vocal cordsGlottis

Characterization of Human Speech

• Frequency: 300-3400 Hz

• The charcaterization of a 20 ms duration sample does not change (excitation, filter system)

• We can find some correlations between two consecutive 20 ms samples

• Using a long time analysis we can predict the next pattern (differencial coding, eg. DPCM)

•Our brain can correct a lot of „errors”!

Classification of Speech Encoders

Waveform encoder: describes the analog signals in the time and frequency domain

Vocoders: describes the rules of phonation (how we generate the voice)

Hybrid encoders: combines good properties of two coders (low rate, good quality)

Subjective MOS

MOS: Mean Opinion Score

Theory of Waveform Encoders

Sampling(sampling frequency)Quantization(number of levels)Coding(length of codeword)

Benefits: good voice qualityDisadvantage: big rate (PCM: 64 kbit/s)

Theory of Vocoders

Benefit: low rate (pár kb/s)Disadvantage: bad quality

A, E, O, I etc.vowels

B, F, P, M etc.consonants

Excitation

Voiced - Vowels

Typical amplitude/time signal

Typical Power Spectrum Density/

Frequency signalPitch frequency: 50-500 Hz

Voiceless - Consonants

Typical amplitude/time signal (white noise)

Typical Power Spectrum Density/

Frequency signal

Sub-band Coding

•Baseband is divided into sub-bands regarding PSD

•Each sub-bands are encoded separately

•Benefits: low noise sensitivity (different sampling frequencies and codeword length are used)

•Bitrate: 16-32 kbit/s

•Receiver: sub-bands are superimposed (added)

Speech Codecs MOS

AMR: Adaptive Multirate Codec(3G és 4G mobile)

G723: VoIP (mobile)

G728: LD-CELP, Low-Delay code excited linear prediction

G729: CS-ACELP, Coding of Speech, Adaptive Code Excited Linear Prediction (VoIP)

GSM FR and EFR: Full rate, Enhanced Full Rate

Internet Low Bitrate Codec (iLBC)

Theory of Hybrid Encoder/Decoder

Analysis by Synthesis (AbS)

Excitation: MPE and RPE

MPE: Multiple Pulse Excitation

RPE: Regular Pulse Excitation, GSM-13kbit/s

MPE: aplitudes and phases are encodedRPE: only amplitudes is encoded

RPE

MPE

Excitation: CELP (Code Excited Linear Prediction)

•Code table: 1024 excitation samples•Only the index of table is sent!•Bitrate: 4,8-16 kbit/s

Excitation Table

Excitation Table

Index

Codec Rate

Codec rate:tin-tout

Typical: 50-100 msFast: G728 CELP (Do-CELP) codec, 2-5 ms

Encoder

Decoder

tin

tout

GSM Speech Codec

Bitrate: 260 bit/20ms, 13kbit/s

Waveform encoder

Voice activity detector

Voice Activity Detector (VAD)

Discontinous Transmission (DTX) and SID Frame

SID: Silence Indicator

Frequency

Fletcher-Munson Curves (sensitivity of ears)

Equal-loudness contours for the human ear

Phon: Unit of relative loudness level for pure tones

Color Sensitivity of Eyes

Luminosity sensitivity is 3 x more faster then color sensitivity

Compression of Source Information

Lossless compression: small part of information is lost

Lossy compression: huge part of information is lost

Compression Rate: original information/compressed information

image: lossless : .raw lossy: .jpeg (eg. 25/1)

Error Correction (general)

Noisy channel Demodulator

Received bit sequence (BER)

Error detectionError detection

procedures

Error correction

Automatic Repeat Request (ARQ)

Forward Error Coding (FEC)

Error Detection I. (repetition)

• We use small blocks• Each block is sent repeatedly• Efficient, when we have only one bit error

1011 1011 1011 1010 1011 1011Sender Receiver

1011 1011 1011 1010 1011 1010

• Even or odd parity• Even: we add „1” parity if the number of „1”s is

odd, otherwise parity is „0”.• Odd: we add „1” parity, if the number of „1”s is

even• Applications: mikrocontollers, data bus, data

storage

10110010

00110010

Even parity

Odd parity

Sender Receiver

10110110

00110110

Error Detection II. (parity)

Error Detection III. (polarity change)

10110010

01001101

Sender Receiver

10110010

01011101

• We send the original and negated bit sequency• If the bit polarity is the same at the receiver -

error• Disadvantege: big redundancy• Application: multicarrier systems

01011101

10100010

Receiver

Error Detection IV. (Cyclic Redundancy Check -CRC)

• A polimon is generated from bit sequence (data) • Generator polinom is the same (known) at sender and

receiver• CRC= data polinom/generator polinom• Application: mobile, RFID, Bluetooth, Ethetnet..etc

10110010101001Sender Receiver

CRC 10100010101001 CRC

polinom

generator polinom generator polinom

division

polinom

division CRC

?

MSB LSB

Error Detection V. (CheckSum)

• The message is segmented into parts• We binary add the parts (CS)• The message and CS are sent• We binary add the parts again at receiver side• If two CSs are not equal -error

Error Detection VI. (Hamming coding)

Eg: 4/7 Hamming code …….4 data bits – 3 parity bits d=2 error can be detected, but only one is correctedNumber of parity bits = largest exponent (23)

Sender

Receiver

Code i Code i

Hamming distance

calculator

The Hamming distance of the code words is greater or equal d+1

Hamming coding (even parity)

7653 p1 p2 p3

0001 1 1 00010 1 0 10011 0 1 10100 0 1 1

DATA PARITY

p1: (3,5,7) – 20=1P2: (3,6,7) – 21=2P3: (5,6,7) – 22=4

7 6 5 4 3 2 1HD:4

HD:3

HD:3

Error Correction(ARQ:Automatic Repeat Request)

Four types:

•Stop and Wait

•Go-Back-N

•Selective

•Hybrid

Stop and Wait ARQ

t1

t2

t1

After time out the sender repeats the transmission

Problem: receiver does not know whether it received the repeated or next data blockSolution: 0 and 1 flag bits in consecutive data blocks

Data block

Data block

Data block

Sender

Sender

Receiver

Receiver

Go-Back-N ARQ1. We define a sliding time window2. The packets are sent one by one (without ACKs) in this window3. If the first ACK is detected window slides.4. If not, all packets are sent again

Selective ARQ

Message

Long message – one ACKAdvantage: if no errors – fast transmissionDisadvantage: If errors – the message is sent again – slow transmission

Small block – more ACKsAdvantage: we repeate only those blocks where errors are detected Disadvantage: slow ( ACK)

Hibrid ARQ (Chase combining)

NACK is sent back, receiver does not drop faulty packetsUsing an errors correction procedure we can repair the data from faulty packets

Hibrid ARQ (Incremental redundancy)

A FEC (Forward Error Coding) is used at sender sideThe redundancy bits are sent after NACK

Forward Error Coding: FEC

Two main types:

•Convolutional coding

•Block coding

We add redundancy bits which help correct data at the receiver

Convolutional Coding (theory example)

(n,k,m) coderD laches and modulo2 adders• n: number of outputs• k: number of inputs• m: number of D laches• number of input state : (m-1)2

(3,1,3) coder

Generator polinomsG1(1,1,1)G2 (0,1,1)G3(1,0,1)

Types of Convolutional Coders (Mother coder and „punctured” coder)

Mother coder:1/n

„Punctured” coder: k/n

eg: 2/3 coder

Systematic coder

•We send original data (V3,V2,V1 and redundancy (v4)

Non systematic coder

•We do not send original data•We send only convolutional coded info

We code „1”(find the error!)

Output:11111011

1011 vs. 1 1 1 1 0 1 1 1 0 1 0 1 1 1 ?

But we use convolution !

Input bits Respones (coded data)1 1 1 1 1 1 0 1 10 0 0 0 0 0 0 0 0 (two bits shift)1 1 1 1 1 1 0 1 11 1 1 1 1 1 0 1 1 -----------------------------------------------------------------------------Kódolt info 1 1 1 1 0 1 1 1 0 1 0 1 1 1

We use modulo2 adder for outputs!

Shift: L-1 (L: number of inner D laches)

Operation table

State graphInfo: 1011Coded info:1 1 1 1 0 1 1 1 0 1 0 1 1 1

Tree structure

Info: 1011Coded info:1 1 1 1 0 1 1 1 0 1 0 1 1 1

Trellis Graph

Coding with Trellis Graph

Info: 1011Coded info:1 1 1 1 0 1 1 1 0 1 0 1 1 1

Error Correction – Fano algorythm I.

Info: 1011Coded info: 11 11 01 11 01 01 11 - Detected info: 01 11 01 11 01 01 11

Error Correction – Fano algorythm II.

Error Correction – Fano algorythm III.

Error Correction – Viterbi algorythm I.

Hamming metricWe use Hamming metric

Error Correction – Viterbi algorythm II.

Error Correction – Viterbi algorythm III.

Error Correction – Viterbi algorythm IV.

Error Correction – Viterbi algorythm V.

Error Correction – Viterbi algorythm VI.

Error Correction – Block Coding (general)

G= 1+ X11+X13+X14+X16

XOR Gates

LFSR: Linear Feedback Shift Register

Error Correction – Block Coding (GSM Speech)

Data50 bits Block

code bitsXOR gate

SW-ON – coding (50 clocks)SW-OFF – Reading Data from D laches