www.final-yearprojects.co.

cc

Speech Compression
using
GSM RPE-LTP
Faiza Nawaz
Bisma Hashmi
Mehrin Kiani
Introduction to GSM
 The Global System for Mobile Communications is the most
popular standard for mobile phones in the world.

 GSM service is used by over 2 billion people across more than

212 countries and territories.

 The ubiquity of the GSM standard makes international roaming

very common between mobile phone operators.

 GSM differs significantly from its predecessors in that both

signaling and speech channels are Digital call quality.
(so it is considered a second generation (2G) mobile phone
system.)

www.final-yearprojects.co.cc 2
Architecture Of GSM

www.final-yearprojects.co.cc 3
What is Speech?

 Speech Generation:

www.final-yearprojects.co.cc 4
GSM 6.10 Vocoder

 Key principle: mathematical modeling of the human vocal tract,

leading to an efficient compression method for transmitting
speech.

 A vocoder (combination of voice and coder) is used to describe

GSM systems tailored for the compression of speech.

 The sampling rate is 8000 sample/s leading to an average bit

rate for the encoded bit stream of 13 K bit/s

www.final-yearprojects.co.cc 5
GSM 6.10 Vocoder
 Coding scheme used by GSM 6.10 Vocoder is the Regular Pulse
Excitation - Long Term prediction - Linear Predictive Coder
(RPE-LTP)

 Vocoder sends three kinds of information to the receiver:

 Voiced or unvoiced signal
 (If it is voiced) The period of the excitation signal
 The parameters of the prediction filter.

www.final-yearprojects.co.cc 6
Linear Predictive Coder (LPC)
 LPC algorithm assumes that each speech sample is a linear
combination of previous samples.

 Speech is sampled, stored and analyzed.

 Coefficients calculated from the sample are transmitted and

processed in the receiver.

 Receiver accurately processes and categorizes voiced and

unvoiced sounds.

www.final-yearprojects.co.cc 7
Residual Pulse Excited (RPE) Coder
 Determines if the signal is voiced or unvoiced

 Determines the period for voiced sounds, encodes periodicity and

transmits the coefficient

 When the signal changes from voiced to unvoiced, RPE transmits a

code that stops the receiver from generating periodic pulses

 Starts generating random pulses to correspond to the noise like

nature of unvoiced

www.final-yearprojects.co.cc 8
GSM Compression Technologies

 Four compression technologies are:

 Full Rate
 Enhanced Full Rate (EFR)
 Adaptive Multi-Rate (AMR)
 Half Rate

www.final-yearprojects.co.cc 9
GSM Full Rate Vocoder Using RPE-LTP

 Described as an RPE-LTP linear predictive coder.

 Models the human vocal tract as a series of cylinders of

different widths.

 By forcing air through these cylinders, speech sounds

can be generated— the LPC coder models this with a
set of simultaneous equations.

www.final-yearprojects.co.cc 10
GSM Full Rate Vocoder Using RPE-LTP
(…contd)
 The input data to the RPE-LTP coder is 20ms of speech
composed of 160 samples, each with 13bit resolution.

 The data is first passed through a pre-emphasis filter:

 Enhances high-frequency components of the signal. (better
transmission efficiency.)
 Also removes any offset on the signal. (Simplifies computation.)

www.final-yearprojects.co.cc 11
LPC Speech Generation
 The model of speech generation can be thought of as air passing
through a set of different size cylinders.

www.final-yearprojects.co.cc 12
Short Term Analysis Stage
 Uses autocorrelation to calculate a set of eight reflection
coefficients.

 Schur recursion is used to efficiently solve the set of

equations resulting from it.

 The parameters are then converted into log-area ratios

(LARs) -- that allow better quantizing in a smaller
number of bits — the first eight parameters of the
transmission stream.

www.final-yearprojects.co.cc 13
Short Term Analysis Stage (…contd)
 The coded LARs is then decoded back to coefficients
and used to filter the input samples.

 The reason for decoding the LARs is to ensure that the

encoder uses the same information available at the
decoder to perform the filtering.

 An array of weights lpc[P] is computed such that

s[n] ~ lpc[0]*s[n--1]+lpc[1]*s[n--2]+_+lpc[P--1]*s[n--P]
(P is usually between 8 and 14, GSM uses 8.)

www.final-yearprojects.co.cc 14
Long Term Prediction Stage

 The 160 samples are split into 4 sub-windows of 40

samples each.

www.final-yearprojects.co.cc 15
Long Term Prediction Stage (…
contd)
 The long-term predictor produces two parameters for
each sub window: the lag and the gain.

 The LTP lag describes the source of the copy in time.

 The LTP gain describes the scaling factor.

www.final-yearprojects.co.cc 16
Calculating Lag and Gain

 LAG:
Compute resemblance by correlation.
correlation of x[n] and y[n] =
Sum of products x[n]*y[n-lag]
 GAIN:
Maximum correlation divided by the energy of the
reconstructed short-term residual signal.

www.final-yearprojects.co.cc 17
Residual Pulse Encoding

 To remove the long-term predictable signal from

its input, the algorithm then subtracts the scaled
40 samples.

 The residual signal is either weak or random and

consequently cheaper to encode and transmit.

www.final-yearprojects.co.cc 18
Residual Signal(…contd)
 The algorithm down-samples by a factor of three,
discarding two out of three sample values.

 Results in four evenly spaced 13-value subsequences to

choose from, starting with samples 1, 2, 3, and 4.

 The algorithm picks the sequence with the most energy.

 That leaves us with 13 3-bit sample values and a 6-bit

scaling factor that turns the PCM encoding into an
APCM

www.final-yearprojects.co.cc 19
Speech Decoder
 Decoder consists of three parts

 RPE Decoding

 LTP synthesis filter

 LPC short term synthesis filter

www.final-yearprojects.co.cc 20
Speech Decoder(…contd)

www.final-yearprojects.co.cc 21
Speech Decoder (…contd)
 Algorithm multiplies the 13 3-bit samples by the scaling factor and
expands them back into 40 samples, zero-padding the gaps

 Resulting residual pulse is fed to the long-term synthesis filter

 40-sample segment is cut from the old estimated short-term residual

signal, scaled by the LTP gain and added to the incoming pulse

 Estimated short-term residual signal passes through the short-term

synthesis filter whose reflection coefficients are calculated by the
LPC module

 Noise from the excited long-term synthesis filter passes through the
tubes of the simulated vocal tract--and emerges as speech

www.final-yearprojects.co.cc 22
QUESTIONS ???

www.final-yearprojects.co.cc 23