Voice Coding in 3G Networks

Tommi Koistinen
Signal Processing Systems
Nokia Networks
[email protected]

Abstract signal is already band limited to 300-3400 Hz there

is no point of using 24-bit converter. Commonly 13
The 3G networks will introduce several new bits per sample is seen to be a practical value for
additions to the basic speech service. The adaptive restricted voice band quantisation. The uniform
wideband speech codec will enhance the quantisation however is not the most efficient
naturalness of speech and the transcoder free quantisation method.
operation will remove unnecessary encodings that
otherwise would degrade the speech quality. The The main idea behind the G.711 standard is to use
speech processing on network side in 3GPP a logarithmic quantizer which results the same
reference architecture model is focused around two signal-to-noise ratio (SNR) with only 8 bits per
network elements, namely the Media Gateway sample compared to original 13 bits per sample.
(MG), and the Media Resource Functions (MRF)
unit. However, as the speech applications utilize the This is achieved by allocating more quantisation
network more or less in transparent end-to-end steps to lower amplitude levels that in fact are the
mode the characteristics and speech enhancement most important to perceived overall speech quality.
capabilities of mobile terminals will finally determine The drawback is that the logarithmic scale will
the perceived overall speech quality. result a reduced SNR in the area of high-powered
input signals but happily the effect of this is
insignificant with speech signals.

1 Introduction As a result we can multiply 8000 samples per

second (that came from the sampling theorem) with
Voice compression techniques have been utilized in 8 bits per sample (that resulted from the logarithmic
digital telecommunication networks for decades quantisation) to get the final bit stream of 64 kbit/s.
(G.711 standard [1] dates back to 1972). The G.711
standard presents a coding technique that operates The compression ratio of G.711 standard can be
at rate of 64 kbit/s and is widely used in all digital seen to be 1.625:1 (13:8). And all compression is
switched telephone networks. But where does the usually good. To transfer more telephone calls with
exact rate of 64 kbit/s come ? less transmission equipment means money for the
operator and this has resulted that several more
The most essential frequency range for the human advanced compression techniques have been
speech production system (that is the glottis and developed.
the vocal tract) and for the auditory system
happens to be between 300-3400 Hz. As the Speech coding techniques in general can be
sampling theorem says; to reproduce the original separated to waveform coders (e.g. G.711, G.726,
signal after sampling we must use a sampling rate G.722) and to analysis-by-synthesis type of coders
that is double the desired frequency band. If (e.g. G.723, G.729, GSM FR). The waveform
sampling rate is less the reproduced signal will be coders operate in time domain and they are based
distorted by image frequencies of the original on sample-by-sample approach that utilizes the
signal. Speech in telecommunication networks is correlation between speech samples. Analysis-by-
commonly sampled at 8 kHz to obey this law. synthesis types of coders try to imitate the human
speech production system by a simplified model of
The number of bits per sample that is used to a source (glottis) and a filter (vocal tract) that
quantize the analog signal is a compromise shapes the output speech spectrum on frame basis
between the quantisation noise that is introduced (typically frame size of 10-30 ms is used). A short
and the quality of the original signal. If the input introduction to details of both basic techniques (and
their intermediate versions; hybrids) is presented in 3GPP has scheduled its work to releases of R99,
[2] on pages 270-287. R4 and R5 and so on. In the following the basic
reference architecture model of each release is
The waveform coders are mainly used to compress shortly described emphasizing the voice coding and
speech on transmission links, for example, on PCM user plane issues.
trunks between two switching centers. The
compression ratios range from 2:1 to 4:1 and quite Release 99
high speech quality can be maintained.
The basic architecture of R99 compatible network is
The analysis-by-synthesis types of coders were shown in Figure 1. The IP packet data from UTRAN
mainly introduced together with digital mobile (Universal Terrestrial Radio Access Network, that is
networks (GSM Full Rate codec [3] dates back to basically base stations and Radio Network
1988). As frequency band in the radio interface Controllers (RNC)) goes through Iu-PS interface to
between a mobile terminal and a base station is 3G SGSN. Voice data goes through Iu-CS interface
restricted (and regulated) compression techniques to 3G Mobile Switching Center (MSC) that converts
are a meaningful way to save money in that the Adaptive Multirate (AMR) coded speech to
interface. A typical full rate channel (16 kbps) G.711 format and vice versa for the PSTN network.
utilizes a compression rate of 4:1. A half rate The circuit switched speech is transferred in packet
channel (8 kbps) is half of that and it operates at mode (ATM/AAL2) from UTRAN (from Radio
compression rate of 8:1. Lossy compression has Network Controller) to 3G MSC but the codec level
always some effects on speech quality and more packet mode speech is not yet originated from the
compression means usually less quality. The G.711 terminal.
standard is common reference point for “real”
speech codecs and e.g. GSM Enhanced Full Rate
codec [4] almost reaches the quality of G.711.
Multimedia
S
SG
GS
SN
N G
GG
GS
SN
N IPnetworks
The frame based handling that is natural to Iu-PS
analysis-by-synthesis coders is also in line with the M
MT
T U
UT
TR
RAN
N
characteristics of packet based transmission H
HL
LR
R
techniques (IP, ATM) that are becoming quite Iu-CS
3
3G
GMS
SC
common not only in core networks (or backbone) T
PSTN/legacy
Tra
ran
ns
sc
co
od
de
err networks
but also as building blocks of radio access
networks.

This article will discuss the voice coding and user Figure 1. 3GPP Reference Architecture of Release
plane issues particularly in 3G networks. The first 99.
chapter presented the basic reasons and means for
speech coding in general. The second chapter will Release R4
review the basic 3G network architecture models.
The most important 3G network elements that The next step that is taken with release 4 (formerly
provide speech related processing are discussed in known as Release 2000) is to separate the
chapters four and five. The sixth chapter will signaling and the user data in Iu-CS interface. The
discuss the issues related to tandeming of speech signaling goes now to MSC Server and the
codecs and finally the seventh chapter will conclude transcoder is separated as a standalone media
the presentation. gateway. Figure 2 presents the R4 architecture with
clear separation to packet side and to circuit
switched side. Media gateway in the PSTN
2 Network Architectures interface converts the AMR coded speech to G.711.
Speech goes in packet mode from UTRAN to PSTN
interface.
This chapter will present the basic 3G network
evolution according to 3GPP (Third Generation
Partnership Project [5]) reference architectures.
 and announcement and conferencing services.
H
HSS
S/C
/CS
SCF Control mechanisms for these functionalities have
usually been proprietary. In 3G networks, all of
Multimedia these functions must be offered by the Media
SG
GSN
N GGSN
N IPnetworks Gateway that is controlled by the Media Gateway
Iu-PS
Controller (MGC) with the standard H.248 control
MT UT
TRAN MGW
W M
MG
GW protocol [7].
Iu-CS PSTN/legacy
user data
networks
An example (and quite full) set of functions that
Iu-CS
control MSCC M
M S
SC Media Gateway could implement is:
S
Serve
err Se
erv
rver
• support for several interfaces (A-interface for
Figure 2. 3GPP Reference Architecture of Release 4. 2G and Iu-interface for 3G) and for several
transmission protocols (ATM, IP, TDM)
The final architecture model, also called as All-IP • support for several codecs including the
network [6], moves also speech to full end-to-end Adaptive Multirate (AMR) codec and future
packet mode. The IP packets that are generated in coming wideband codecs
a mobile terminal go as such either to another IP • electric and acoustic echo cancellation
terminal or to MGW from GGSN. The architecture is • announcement services
presented in Figure 3. A new network entity is also • DTMF and call progress tone generation and
introduced, namely the Multimedia Resource detection
Functions (MRF) unit that implements mainly • support for fax/modem/data protocols
conferencing services for the IP based calls. • support for Tandem Free Operation (TFO) and
Transcoder Free Operation (TrFO)
• bad frame handling
H
HSS/CSC
CF
F • IP protocol handling (RTP/RTCP, encryption,
QoS support)
Multimedia
S
SGSN G
GG
GS
SN IPnetworks
Iu-PS Some functions, especially the conferencing service
MT U
and possible speech enhancement services, are
T UT
TR
RAN MRF
F MG
GW
W
basically thought to be provided by the Multimedia
PSTN/legacy Resource Functions (MRF) unit, but they may
networks optionally be added to Media Gateway
responsibilities.

A lot of signal processing (DSP) power is required

Figure 3. All-IP reference architecture. to provide the Media Gateway’s functions.
Typically, one DSP chip may process 4-16
Of course, the different phases of 3GPP releases channels, and on one processor card there might
may coexist at the same time depending on be 8-32 DSPs which totals 32-512 channels per
operators’ needs. processor card.

4 Media Resource Functions

3 Media Gateway
The Multimedia Resource Functions (MRF) unit
In 2G networks (like GSM) the speech related according to 3GPP standard shall provide the
functionalities have been implemented around the audio/video conferencing services for the All-IP
transcoder unit (TRAU). The basic task of network. The basic requirement is to support
transcoder has been speech encoding and several speech codecs to be able to sum up the
decoding of narrowband codecs like GSM Full Rate conference for each party. As it is impossible for
(FR), Enhanced Full Rate (EFR) or Half Rate (HR) today’s technology to sum up signals in parameter
codecs. Some extra features like noise cancellation domain, all signals must be first decoded for linear
or acoustic echo cancellation are also offered by domain processing. The summed signals are then
2G transcoders. The Mobile Switching Center has encoded again for each party.
then additionally offered tone and DTMF
generators, echo cancellers, fax and modem pools The 3GPP work on MRF entity has not progressed
further than the conferencing requirement.
However, the MRF entity is a natural place also for speech. Also the transcoders must support TFO
other speech enhancement services. It should be feature as they must omit the decoding and pass
remembered that most of the calls in an All-IP encoded parameters as such forward.
network are staying inside the core network and An end-to-end connection (of 16 or 8 kbps) can
they are not going to Media Gateway at all (see now be formed with only one encoding (in
figure 4). originating mobile) and only one decoding (in
receiving mobile). The figures 5 and 6 present the
cases without TFO and with TFO in operation.
MT
T U
UTRA
AN
N MRF IP
IPte
term
rmin
ina
al

Multimedia
S
SG
GSN G
GGSN IPnetworks
MSC PSTN MSC
C
64kbps
MT
T U
UTRA
AN
N Tran
ranscoder Tra
ransco
oder
64↔ 16 64↔ 16
Figure 4. MRF unit as a network side speech
enhancement server.
BSS BS
SS
Calls between mobile IP terminals are transferred in
M
MS MS
coded format end-to-end and if any speech
enhancement services are desired to be provided
on the network side, the MRF entity could do the
necessary operations (as it already has to support Figure 5. No Tandem Free Operation.
all coding formats for the conferencing service).
The other option is that all speech enhancement
services shall be provided by mobile terminals.
MSC PSTN MSC
48(16) kbps
A set of speech enhancements that the MRF entity
could provide is: Transcode
er Tra
ranscode
er
16↔ 16 16↔ 16

• Noise suppression
• Gain (volume) control B
BSS
S B
BSS
S
• Acoustic echo cancellation
M
MS MS
S
It should also be mentioned that the Media
Gateway and the Multimedia Resource Functions
unit are logical entities only and physically they may Figure 6. Tandem Free Operation is utilised.
co-locate in the same device.
TFO is based on inband procedures that means
5 Tandem Avoidance that no outband signaling is used to form a TFO
connection. In practice, the TFO connection
establishment starts with a negotiation phase where
5.1 Tandem Free Operation (TFO) certain TFO protocol messages are exchanged
between transcoders to agree on the used codecs.
Every time voice is encoded or decoded the speech
If the other end doesn’t support TFO it will not
quality will degrade a little bit. Thus, as few
acknowledge the negotiation and also the TFO
conversion as possible are desired. The basic 2G
capable transcoder will start to encode and decode
mobile-to-mobile call suffers from tandem coding
the 64 kbps as in figure 5.
that means that separate speech coding happens
in both radio interfaces and between the
transcoders voice goes in 64 kbps G.711 format. In 5.2 Transcoder Free Operation (TrFO)
general two encodings in clear speech conditions is
no problem but more than two encodings especially For the 3G networks a slightly different approach is
in bad line conditions cause severe degradations. taken considering tandem avoidance. Firstly,
outband signaling is used for codec negotiation and
To overcome this kind of quality problem ETSI has if codecs match there is no need for the
specified so called Tandem Free Operation (TFO) transcoders at all. Operation is called as
[8] that establishes a sub channel (of 16 or 8 kbps) Transcoder Free Operation (TrFO) [9].
inside the 64 kbps G.711 stream for the encoded
TrFO is relevant mainly for the MSC Server concept during silence only silence description (SID) frames
and for intersystem compatibility as in the final All- are periodically sent to other end. All modes
IP network calls are by nature of TrFO type. In operate on 20 ms frame basis.
figure 7 is presented a basic call where outband
signaling travels from MSC Server to another until
the whole link is negotiated. If a common codec can Codec mode Source codec bit-rate
be agreed no transcoding resources are reserved AMR_12.20 12.20 kbit/s FR
from the intermediate media gateways. AMR_10.20 10.20 kbit/s FR
AMR_7.95 7.95 kbit/s FR / HR
AMR EFR! AMR_7.40 7.40 kbit/s FR / HR
M
MT
T U
UT
TR
RA
AN
N M
MG
GW
W M
MG
GW
W G
GS
SMBS
SS
S AMR_6.70 6.70 kbit/s FR / HR
AMR?
AMR_5.90 5.90 kbit/s FR / HR
M
MSSC
C M
MSSC
C
PSTN/legacy AMR_5.15 5.15 kbit/s FR / HR
S networks
Se
erv
rve
err S
Se
erv
rve
err
AMR?
AMR_4.75 4.75 kbit/s FR / HR
AMR_SID 1.80 kbit/s FR / HR
Figure 7. A basic TrFO call.
Table 1. 8+1 different AMR modes.

The choice between the full rate and the half rate
4 Adaptive Speech Coding channel mode can be made off-line based on the
capacity requirements of the operator. The
The traditional GSM speech codecs operate in the selection of the codec mode happens continuously
radio interface at a fixed source rate with a fixed by the radio resource management. Basically, as a
level of error protection (e.g. Full Rate codec with lower AMR mode is selected, more bits from the
framing overhead consumes 16 kbps and error gross bit rate are freed for the channel coding and
protection adds 6.8 kbps resulting a 22.8 kbps error protection. Even that we use a very low codec
gross bit rate over the air). The codec itself do not bit rate the high error protection keeps the overall
have means (except bad frame handling speech quality sufficiently high. The figure 8 shows
mechanism) to adapt to changing radio conditions. reasoning for the mode selection. To follow the
For this reason, ETSI (and later 3GPP) has asked optimum quality curve (MOS=Mean Opinion Score
for new adaptive coding schemes that could select of speech quality) against decreasing signal-to-
the optimum channel mode (full rate or half rate) noise ratio (C/I) the AMR mode that is used must
and the optimum codec mode (speech rates) based be changed accordingly.
on the radio conditions. As a result, the Adaptive
Multirate (AMR) codec [10,11] has now been
standardized as an additional codec for the GSM M
system and as the only mandatory codec (thus far) O
for the 3G system. Two most important design S
targets for the AMR codec were: Mode 1
Mode 2
• improved speech quality in both half-rate and Mode 3
full-rate modes by means of codec mode
adaptation i.e. varying the balance between
speech and channel coding for the same gross
bit-rate.
C/I

• ability to trade speech quality and capacity

smoothly and flexibly by a combination of
Figure 8. Different AMR modes have different
channel and codec mode adaptation; this can
quality curves.
be controlled by the network operator on a cell
by cell basis.
It should be however noted that in the 3G radio
interface the power control mechanism (fast power
The AMR codec consist of 2 channel modes (full
control and outer loop power control) is used to
rate (FR) and half rate (HR)) and 8 codec modes
keep the optimum speech quality by adjusting the
that are presented in table 1. The ninth mode is for
transmit power of a mobile terminal and the base
discontinuous transmission (DTX) meaning that
station. The adaptiveness of AMR in fact doesn’t Parameter 1st 2nd 3rd 4th Total
bring such benefits for 3G as it does for 2G radio 2 LSP sets 38
interface.
Pitch delay 9 6 9 6 30
Principles of the AMR encoder Pitch gain 4 4 4 4 16
Fixed code 35 35 35 35 140
The AMR codec is based on the Code-Excited Fixed gain 5 5 5 5 20
Linear Predictive (CELP) coding model that imitates Total 244
the glottis and the vocal tract by an excitation signal
and a linear prediction synthesis filter (Figure 9).
Table 2. Encoder output.

RTP payload specification for AMR codec

adaptive codebook gp
In the 3GPP Release 99 architecture the AMR
v(n) codec payload is packed in the Radio Network
fixed
+
u(n) 1 ^
s(n) post-filtering
^
s'(n) Controller in IuUP protocol frames [12] that are
codebook A(z)
carried as such to transcoder in 3G MSC. The
gc
c(n) LP synthesis specified frame format for AMR codec is restricted
to Iu interface.

In the All-IP model (figure 3) the AMR payload data

travels all the way from the mobile terminal through
Figure 9. The CELP model.
UTRAN and the core network either to media
gateway or another IP terminal. The GGSN will
The excitation signal at the input of the short-term
output the application level protocols, that in this
LP synthesis filter is constructed by adding two
case, are the RTP (Real-time Transport Protocol)
excitation vectors from adaptive and fixed
frames carrying the AMR payloads. So, concerning
codebooks. The speech is synthesized by feeding
IP Telephony the RTP payload specification for
the two properly chosen vectors from these
AMR codec [13] has grown in importance as AMR
codebooks through the short-term synthesis filter.
is the codec that should converge the traditional IP
The optimum excitation sequence in a codebook is
Telephony with the mobile IP Telephony. The RTP
chosen using an analysis-by-synthesis search
for AMR specification includes the following extra
procedure.
features:
The AMR coder operates on speech frames of 20
• codec mode request procedure
ms corresponding to 160 samples at the sampling
frequency of 8 000 sample/s. At each 160 speech • robust sorting of payload bits
samples, the speech signal is analysed to extract • bad frame indication
the parameters of the CELP model (LP filter • compound payloads
coefficients, adaptive and fixed codebooks' indices • CRC calculation
and gains). These parameters are encoded and
transmitted. At the decoder, these parameters are The specification is still under finalisation in IETF.
decoded and speech is synthesized by filtering the
reconstructed excitation signal through the LP 5 Wideband Speech Coding
synthesis filter.
The 300-3400Hz speech band frequency range has
Table 2 shows the resulting parameters from the
been used for decades in all telephony
encoder operating in 12.2 kbps mode. The LP
applications. As the range is heavily restricted all
analysis is performed twice per 20ms frame
non-speech signals, like music, are degraded badly
resulting 2 sets of line spectrum pairs (LSP). The
when forced to go through this narrow frequency
adaptive codebook (pitch delay and gain) and the
pipe. Even speech contains plenty of information
fixed codebook are found for 4 subframes of 5 ms
above 3400 Hz that affects the naturalness of
each. Total number of bits is 244 per frame.
speech.

Basically, the existing terminals that conform to this

traditional frequency band have been one barrier in
front of wideband speech. Second reason has been
 that more bandwidth is needed to transfer the Codec mode Source codec bit-rate
highest quality wideband signals. AMR-WB_23.85 23.80 kbit/s
AMR-WB_23.05 23.05 kbit/s
However, as the difference in quality between
AMR-WB_19.85 19.85 kbit/s
narrowband and wideband speech is so clear it is
evitable that more wideband applications will be AMR-WB_18.25 18.25 kbit/s
introduced in the near future. Wideband speech AMR-WB_15.85 15.85 kbit/s
coding can easily be seen as the next fundamental AMR-WB_14.25 14.25kbit/s
improvement in speech quality for mobile AMR-WB_12.65 12.65 kbit/s
telecommunication systems. 3GPP has understood
this and wideband AMR specifications are already AMR-WB_8.85 8.85 kbit/s
getting ready. AMR-WB_6.6 6.6 kbit/s
AMR-WB_SID 1.75 kbit/s
The principles of wideband AMR [14] are copied
from the narrowband AMR. The frequency band, as Table 3. 9 Different AMR-WB modes.
a difference, is extended in both directions, and it is
now from 50 Hz to 7000 Hz. The resulting speech
quality exceeds the wireline quality of narrowband 6 Conclusion
G.711. In figure 10 is shown an illustrive graph on
speech quality comparison of EFR, AMR-NB and Packet data services have been advertised to be
AMR-WB in 16kbps full-rate channel [15]. the major application of future 3G networks.
However, also the voice services are strongly
enhanced with new wideband codecs that can
adapt to network conditions. Also the transcoder
AMR-WB
free operation, and the new speech enhancement
Subjectivespeech quality

Excellent
AMR-NB
services will make speech quality better, even to
Verygood level never experienced before.
EFR
Good This article has mainly focused on the application
level. Good network conditions (low delay, no lost
Poor packets due congestion) are a starting point also for
superior application level speech quality. Media
Unacceptable gateways shall support the network level QoS
mechanisms (like DiffServ) that are used to
Error-free 13 10 7 4 optimize and prioritise the real-time and the non-
real-time traffic (see for example [16]).
Carrier-to-interfaceratio (dB)

In the past, speech service has been closely tied on

Figure 10. AMR-WB vs. AMR-NB and EFR. technical level to providing network. Within All-IP
networks also speech service will be lifted more
As can be seen, in error-free conditions the AMR- and more up to user-level. End-to-end user
WB is superior over AMR-NB or EFR (that is the applications will not even see the underlying
highest quality GSM codec at the moment). Even in transport network and the overall speech quality
very bad conditions AMR-WB can maintain high that is perceived will heavily depend on the
quality far above fixed rate GSM codecs. The nine characteristics and features of the All-IP terminals.
modes of AMR-WB (plus one mode for DTX) are
presented in in table 3. As also the speech service will include more
choices of used codecs, used bandwidth and used
AMR-WB is specified for GSM full rate radio traffic speech enhancements there shall be opportunity to
channel, for future GSM EDGE (GERAN) and for differentiate the pricing of these features. The user
the 3G (UTRAN) radio channel. The 3GPP may in the future have means to select the speech
specifications for a wideband AMR codec (AMR- quality that he or she is willing to pay.
WB) are expected to be finalized in March 2001.
References
[1] ITU-T G.711; Pulse Code Modulation (PCM)
of Voice Frequencies. 1972.

[2] Hersent O, Gurle D, Petit J-P. IP Telephony.

Packet-based multimedia communications
system. Addison Wesley, 2000.

[3] GSM 06.10; Full Rate Speech; Transcoding.

[4] GSM 06.60; Enhanced Full Rate Speech;

Transcoding.

[5] Third Generation Partnership Project (3GPP)

www.3gpp.org

[6] 3GPP/TR 23.922; Architecture for an All-IP

network, v1.0.0, October 1999.

[7] ITU-T H.248; Gateway Control Protocol,

June 2000.

[8] GSM 08.62; Inband Tandem Free Operation

(TFO) of Speech Codecs; Service
Description; Stage, v8.0.1, August 2000.

[9] 3GPP/TS 23.153; Out of Band Transcoder

Control – Stage 2, v2.0.3, October 2000.

[10] 3GPP/TS 26.071; AMR Speech Codec;

General Description, v3.0.1, August 1999.

[11] 3GPP/TR 26.975; Performance

Characterization of AMR Speech Codec,
v1.1.0, January 2000.

[12] 3GPP/TS 25.415; UTRAN Iu Interface User

Plane Protocols, v3.5.0, December 2000.

[13] IETF Internet Draft: RTP Payload Format and

File Storage Format for AMR Audio, v0.5,
February 2001.

[14] 3GPP/TR 26.901; AMR Wideband Speech

Codec; Feasibility Study Report, v4.0.1, April
2000.

[15] Advance – Information from Nokia Research

Center. Number 1, 2001.

[16] Ferguson P, Huston G. Quality of Service;

Delivering QoS on the Internet and in
Corporate Networks. Wiley 1998.