CELLULAR COMMUNICATIONS

5. Speech Coding

Low Bit-rate Voice Coding

Voice is an analogue signal Needed to be transformed in a digital

form (bits) Speech signal is not random=>can be

encoded using fewer bits as compared to random signal

If bits representing 1sec of speech can transferred over wireless channel during 200ms=> can pack 5 signals into the channel

For a handset transmitting less bits is alsoe means longer battery life

Requirement for speech coding Can distort a speech a little bit (lossy)

but should preserve acceptable quality

Shouldn’t be to complex Use less power Use less circuits Reduce delay

Hierarchy of speech coders

Waveform Coders vs. VOCODERS Waveform coders

Approximate any acoustic signal

VOCODERS Based on prior knowledge of the signal Speech signals are very special signals

Speech signals

Not all levels of a speech signal are equally likely High probabilities of very low amplitudes Significant probability of very high

amplitudes Monotonically decreasing probabilities of

amplitudes between these two extremes Speech is predictable

The next value of a speech signals can be predicted with large probability and fair precision from the past samples

Speech in frequency domain Power of high frequency components is

small High frequency components when

present are very important for speech quality

Sampling and quantization

Speech signal is analog, measured at infinitely many time instances and infinitely many possible values

Sampling: measure signal at finite time instances (sampling interval)

Quantization: approximate infinitely many possible values by finite number of possible values (e.g. 8 bits)

Uniform quantization

Divide the range of all possible values into finite number of equal intervals

Assign single quantization value to all values within the interval

Non-uniform quantization

Divide the range of all possible values into finite number of unequal but equally probable intervals

Logarithmic quantization: smaller intervals at low amplitudes

Different weight to low values US: -Law Europe: A-low

-Law

A-law

Adaptive quantization

Adjust to input signal power

Rate-Distortion Theorem

Shannon: There existing a mapping from source waveform to code words such that for given distortion (error) D, R(D) bits per sample is sufficient to restore signal with an average distortion arbitrary close to D

R(D) is called rate distortion formula (achievable low bound)

Scalar quantization does not achieve this bound

Vector quantization

Encode a segment of sampled analog signal (e.g. L samples)

Use codebooks of n vectors Segment all possible

samples of dimension L into areas of equal probability

Very efficient at very low rates( R=0.5 bits per sample)

Learning codebook

LBG: Split areas (double codebook)

Adaptive Differential Pulse Code Modulation

PCM Each sample representing by its amplitude

(8 bits) Standard telephony: 8K samples per

second, 8 bit per sample= 64kbps DPCM

Encode only difference from previous sample

Smaller differences are more often Use less bits to represent smaller

differences(4 bits) but more bits (10 bits) to represent larger differences

DPCM and prediction

ADPCM

Use more complex prediction in a transmitter/receiver to estimate next sample value

Transmitter send only difference between estimation and real value

Lossy codec: transmit approximate differences

Hopefully difference will be small

Frequency Domain Coding of Speech

Divide speech signal into a set of frequency components

Quantize and encode each component separately

Control number of bits/quality allocate to each band

Sub-band coding

Human ear does not detect error at all frequencies equally well

SBC

Vocoders

Model speech signal generation process Transmitter analyze the voice signal

according to assumed model Transmitter sends parameters driveled

from the analysis Receiver synthesize voice based on

received parameters Vocoders are much more complex that

waveform coders but achieve higher economy in a bit rate

Human Vocal Tract

Voice Generation Model

LPC

Advanced codecs

CELP Transmitter/Receiver share common pitch

codebook Search for most suitable pitch code

RELP Transmit model parameters Also transmit Residual(differences) signal

Mean Opinion Score Quality Rating

Codec MOS rating