VoIP Monitoring:

The Codec Challenge

EVS forces VoLTE operators to act

2020
Introduction 3

Audio Codecs Primer 4

(Estimating) Better User Experience 9

Monitoring Multi-Mode Codecs 12

2
Introduction

Voice-over-IP (VoIP) services may encode audio signals

in different ways. A recent study by Voipfuture found
that around a dozen codecs2019are in active use for live
VoLTE, VoWifi, international wholesale, and domestic
VoIP services. The currently dominant codecs are G.711
for fixed voice services and AMR/AMR-WB for mobile
services. However, the codec landscape is changing
and, in both domains, new modern codecs are
becoming more common. 3
Many new codecs offer multiple bitrates, which allow
to adapt voice transmission to the available channel
conditions and bandwidths. Such codec mode changes
can occur on a packet-by-packet basis without any
indication in the SIP signaling. VoIP monitoring needs
to be aware of such mode changes, mainly because the
user experience strongly depends on the bitrate.

This whitepaper provides an overview of codecs for

VoIP communication services and their characteristics.
Based on this, the paper discusses the impact of new
wideband multi-mode codecs on service monitoring
and the resulting requirements on VoIP monitoring
systems.
 Audio Codecs Primer
A codec is software or hardware that implements an algorithm
for encoding and decoding a signal. The coder function encodes

4
signals for transmission or storage, while the decoder function
recreates the signal for playback or editing.

In principle, this definition applies to any form of signal, but the

word codec is most frequently used in the context of audio and
video applications.

More specifically, an audio codec compresses and decompresses

digital audio data according to a given coding format.

The purpose of an audio coder is to convert an audio signal

into a representation with fewer number of bits while retaining
quality. This effectively reduces the bandwidth required for
transmission of audio data.

Hello Hello
Conversion Conversion
Analog/Digital Digital/Analog
Codec Codec
(Encode) (Decode)

Protocol Protocol
Stack Stack
 Codecs are often optimized to encode human speech as
opposed to general audio signals. Human speech and human
hearing have well-known characteristics.

For example, the human ear is capable of hearing frequencies

in the range of 20 Hz to 20,000 Hz while human voice mainly –
not exclusively – uses the frequency range from 300 Hz to 3,400
Hz. In telephony, the latter frequency range is referred to as
narrowband.

By exploiting the properties of human voice and hearing,

narrowband codecs can be used to efficiently transmit speech
signals. This was long state-of-the-art for typical phone
conversations but obviously has some limitations. For example,
inbound call centers often play music to customers waiting to
be served. Anyone who has had to listen to distorted music for
longer periods of time knows that this can be very annoying.

Wideband codecs greatly extend the frequency range from 150

Hz to 7000 Hz for more natural voice transmission. Beyond this,
super-wideband (or ultra-wideband) codecs cover the frequency
range up to 16,000 Hz enabling the transmission of high
definition audio including speech and music.
5
Finally, fullband codecs cover the entire frequency spectrum of
human hearing and beyond. In simple terms, human hearing
cannot distinguish between an original signal and its fullband
encoded version; this is referred to as transparency. The voice
quality delivered by codecs covering wideband and above will
further be called high definition (HD) voice and respective codecs
will further be summarized as HD codecs.

Narrowband Wideband Super Wideband Fullband

300 Hz – 3.4 kHz 150 Hz – 7 kHz 150 Hz – 16 kHz 20 Hz – 22 kHz
Codecs can be characterized in the following ways:

ƒ SAMPLE RATE/BANDWIDTH: The majority of codecs used today sample

the audio signal at a rate of 8,000 Hz. The reason is that the sampling
frequency must be at least twice the highest component of the voice
frequency (3,400 Hz) for effective reconstruction of the signal. Wideband
codecs, such as G.722, have a sample rate of 16,000 Hz.
ƒ BITS PER SAMPLE: The value range of a sample, typically 8 or 16 bit
allowing for 256 respectively 65536 sample values.
ƒ BITRATE: The nominal bitrate of a codec is determined by the size of the
compressed audio signal. The bitrate is constant for most codecs although
there are exceptions from this rule. Codecs typically process collections
of audio samples known as frames. For example, a G.729 frame covers
10 milliseconds (ms) of speech and has a size of 10 bytes, which yields
a bitrate of 8 kbit/s. The required network bandwidth for transmission is
higher than a codec’s bitrate, since headers for application, transport, and
network protocols need to be added.
ƒ LATENCY: Codecs also differ in the amount of algorithmic delay or latency
they introduce. This paper only considers so-called conversational codecs,
i.e. codecs with a latency that is sufficiently low to facilitate bi-directional
conversations.
ƒ COMPLEXITY: Codecs have different requirements in terms of the process-
ing power needed to encode audio. Codecs with a high complexity are
less suitable for battery-powered devices, such as mobile phones.
ƒ SPECIAL FEATURES: Modern codec often include built-in functionality,
such as voice activity detection, silence suppression/discontinued transmis-
sion, comfort-noise generation and the ability to automatically adapt to
6
current network conditions. Most modern codecs implement functionality
that goes beyond plain encoding and decoding of audio. This includes
error concealment mechanisms to mitigate the impact of packet loss,
voice activity detection, silence suppression/discontinued transmission,
comfort-noise generation and the ability to automatically adapt to
current network conditions.
ƒ USER EXPERIENCE: The codec quality in terms of the user experience
under optimal conditions varies considerably, mainly depending on the co-
decs’ bitrate and effectiveness. User experience is typically measured using
the Mean Opinion Score (MOS) or R-factor.
ƒ COST: Not all codecs are free of charge. Indeed, most modern codecs
require a license.
ƒ SUPPORT: Numerous codecs have been proposed and standardized over
the past decades. Few are widely supported by telecommunications equip-
ment, such as mobile handsets, IP phones, MGWs, and SBCs. Even fewer
are in active use.

Most codecs relevant to landline and mobile VoIP telephony are standardized
by ITU-T and 3GPP. One notable exception is the Opus codec, which
was introduced by the IETF. The table lists some codecs and their key
characteristics.
The table (next page) is by no means complete and much more could be said
about each individual codec. Please refer to the respective standards for more
detailed information.
 Codec Bandwidth Net Bitrate Usage

AMR NB 4.75 - 12.2 Widely used multi-mode codec in 2G/3G mobile

(Adaptive kbit/s networks
Multi-Rate)

EVS NB, WB, 5.9 - 128 Currently being deployed in 4G mobile networks;
(Enhanced SWB, FB kbit/s mandatory codec for 5G voice services
Voice
Services)

G.711 NB 64 kbit/s Standard codec for fixed line voice services

G.711.1 NB,WB 64, 80, 96 Not widely supported, but used as HD codec by
kbit/s some fixed voice operators

G.722 WB 64 kbit/s Widely supported HD codec for fixed

line services

G.722.1 WB 24, 32 kbit/s Mainly used by videoconferencing systems

G.722.2 WB 6.60, 8.85, Widely used multi-mode codec in 3G/4G mobile

(AMR-WB) 12.65, 14.25, networks
15.85, 18.25,
19.85, 23.05,
23.85 kbit/s
7
G.723.1 NB 5.3, 6.3 Widely supported by fixed network devices, but
kbit/s not widely used

G.726 NB 16, 24, 32, Widely supported by fixed network devices, but
40 kbit/s not widely used

G.729 NB 8 kbit/s Widely supported by fixed network devices and

frequently used

G.729.1 NB 8 kbit/s, Not widely supported

12–32 kbit/s
in 2 kbit/s
steps

iLBC NB 13.33, Widely supported by fixed network devices, but

15.20 kbit/s not widely used

Opus NB, WB, 6–510 kbit/s Widely supported by fixed network devices, but
SWB, FB not widely used Multi-mode audio codec defined
by the IETF in RFC 6716; required by WebRTC
implementations and slowly gaining support by
fixed network devices

NB = narrowband, WB=wideband, SWB=super-wideband, FB=fullband

 Modern HD codecs do not only provide a better user experience
but are also extremely efficient. This can be seen if one divides
a codec’s maximum achievable R-Factor value by its nominal
bitrate. This quality-per-bit ratio has steadily increased over the
years.

4/11/2019 Page 1

Development of Coding Effiency

EVS WB 9.6k

10
0.010
Quality (R-Factor) per kilobit

AMR WB 12.65k

AMR 12.2k

5
0.005

G.726 32k
G.722
G.711

1970 1980 1990 2000 2010 2020

8
1/1

Coding Efficiency: R-factor per Bit

Currently, the industry is in a transition phase and moving

towards HD voice. Advantages typically cited by users are:
ƒ Clearer overall sound quality
ƒ Easier to recognize voices, distinguish confusing sounds and
understand accented speakers
ƒ Reduced listening effort, resulting in increased productivity
and lessened listener fatigue

The biggest driver for this development is certainly the

introduction of AMR-WB and EVS by VoLTE operators.

But also fixed network operators are – at least for on-net calls
– slowly moving to G.722 and G.711.1. Finally, IPX wholesale
services are enabling HD voice continuity in international calls.
(Estimating) Better
User Experience 9
The rise of wideband codecs creates new challenges for monitor-
ing voice quality and user experience. The Mean Opinion Score
(MOS) is considered the key metric for voice quality, yet few are
aware of its many different flavors.

Historically, MOS is a subjective measurement. A set of listen-

ers in a “quiet room” score the quality of a call on a scale of 1
(“bad”) to 5 (“excellent”). The result is an average, i.e. the mean
opinion of all listeners. All monitoring tools that provide a MOS
essentially attempt to estimate the outcome of such an empirical
study.

ITU-T Recommendation P.800.1 specifies terminology to dis-

tinguish different types of MOS. In effect, P.800.1 defines nine
different types, but only two are relevant in the context of VoIP
monitoring:
 ƒ MOSLQE (listening quality estimate), i.e. the MOS provided by
passive (no reference single-ended) VoIP monitoring systems
and
ƒ MOSLQO (objective listening quality), i.e. the MOS calculated
by full reference end-to-end test systems, e.g. based on ITU-T
P.862 (PESQ) or P.863 (POLQA).

The 2006 revision of P.800.1 introduced notation to specify the

reference audio bandwidth, i.e. the MOS scale. Based on this
the letters N, W, S and F for narrowband, wideband, super-
wideband and fullband should be appended to denote the
quality reference. For example, a MOS listening quality estimate
with a wideband reference should be denoted as MOSLQEW.
Unfortunately, this precise notation is rarely used in practice.

Stating the quality reference is however necessary because

narrowband codecs no longer define the user expectations.
With the introduction of wideband, super-wideband and even
fullband codecs user expectations have started to shift. What
was “excellent” before, is now considered mediocre.

Impact of Scale Choice on MOS Value

4/9/2019 Page 1

MOS_NB
5.0
MOS_WB 10
4.5

4.0
Mean Opinion Score (MOS)

93
3.5

MOS_NB 4.41
3.0
MOS_WB 3.69
2.5

2.0

1.5

1.0
0 20 40 60 80 100 120 140
R-factor
1/1

As a result, new MOS scales had to be introduced, which take

the varying user contexts and expectations into account. The
traditional MOS scale is now often referred to as narrowband
MOS (MOSNB). It is complemented by the wideband (MOSWB)
and super-wideband (MOSSWB) scales. Because of this, the same
codec now has different maximum MOS values, depending on
which scale is used. For example, G.711 has a MOSLQE of 4.41 on
the narrowband scale, but only 3.69 on the wideband scale.
An anecdote illustrates why new MOS scales are needed. Some
years ago, a large German bank deployed a new VoIP system for
the entire company. The selected phones supported standard
narrowband G.711 as well as the G.722 wideband codec. It was
therefore decided to use G.722 where possible to benefit from
higher call clarity, i.e.
ƒ G.722 was used for internal calls between the new IP phones;
ƒ G.711 was used for external calls because this was the only
codec supported for the interconnection with the provider.

What initially seemed like a good idea turned out to be proble-

matic as employees started complaining at the bank’s IT help
desk. The users were perceiving ‘bad quality’ when talking with
customers, even though the quality of those calls was perfect
landline toll quality. Apparently, the employees accepted G.722
wideband audio as the new normal and adjusted their view on
G.711. Without objective quality becoming worse, the subjective
quality perception dropped – after a few months it was decided
to use only G.711 to ensure a consistent narrowband user ex-
perience. It can be expected that the introduction of HD codecs
for VoLTE calls will cause the same habituation effect as people
quickly get used to new levels of quality.

The applicable scale depends on the (assumed) expectations of

11
a user group. VoLTE users will quickly become used to wideband
codecs, i.e. their user experience should be judged on the MOSWB
scale. In contrast, the reference quality of fixed line users is still
standard G.711 quality, therefore MOSNB should be applied. Of
course, this is a simplification as user groups are not mutually
exclusive and expectations will generally rise as users are beco-
ming more exposed to HD voice quality. In the current transition
period, the different MOS scales present an irritating source of
confusion to the industry, specifically if values based on different
reference qualities are compared.

Technically, passive monitoring systems calculate the MOSLQE

using the E-Model defined in G.107. The E-Model specifies how
to determine the so-called R-Factor, which is then transformed
into a MOS value. Originally, the R-Factor scale went from 0 to
100, with an R-Factor of 93 for G.711. For wideband codecs
this scale was extended to 129. Two different transformation
rules provide a mapping from the R-Factor to either a MOSNB
or MOSWB. In June 2019 Recommendation ITU-T G.107.2
introduced the latest update to the E-Model by extending the
R-Factor scale to 148 and providing a transformation rule for
fullband MOS.
 Monitoring
Multi-Mode Codecs 12
Looking more closely at the Enhanced Voice Service (EVS) codec
illustrates the challenge of monitoring user experience when
using multi-mode audio codecs. EVS supports all four bandwidth
modes NB, WB, SWB and FB and each bandwidth offers a set of
bitrate modes.

kbit/sec 5.9 7.2 8 9.6 13.2 16.4 24.4 32 48 64 96 128

NB x x x x x x x - - - - -
WB x x x x x x x x x x x x
SWB - - - x x x x x x x x x
FB - - - - x x x x x x x x

The modes are controlled in different ways and can be switched

every 20 ms.
CONTROL OF MODES VIA SDP
The codec to be used for a call is negotiated at call setup time via the Ses-
sion Description Protocol (SDP). For EVS this entails that the call parties also
agree on a bandwidth and bitrate (or range thereof). This is done via the
codec-specific parameters ‘bw’ and ‘br’ as shown in the following examples:

bw=nb-swb;
br=16.4;
br=13.2-24.4;

Ranges may initially be specified by the caller. The callee may then accept or
not accept the EVS codec as a whole. If he accepts, he shall only answer with
ranges within the constraints given by the caller, as in this example:

Caller SDP: bw=nb-wb;

Callee SDP: bw=wb;

Here, the call parties agree to restrict bandwidth to wideband mode.

CONTROL VIA CMR BYTE

The Codec Mode Request (CMR) byte within the EVS codec payload header
can be used to request a certain bandwidth/bitrate mode from the remo-
13
te call party. The requested modes shall be within the range agreed upon
via SDP. Once a CMR has been received, the receiving party should use this
mode, and may use lower modes within the SDP constraints.

CONTROL BY ENCODER
Finally, the EVS encoder performs bandwidth detection on the input signal
to apply a bandwidth decision logic. This means that the encoder may deci-
de to encode the audio in a lower bandwidth than determined by the input
sampling rate. For example, assume the input sampling frequency is 32 kHz.
If the bandwidth detection logic determines that there is no “energetically
meaningful” spectral content above 8 kHz, then the codec is operated in the
WB mode.

Interestingly, this codec decision does not seem to be constrained by the re-
sult of the SDP negotiation. Voipfuture analysed live VoLTE traffic and found
many examples of calls where SDP negotiations settled on EVS WB as band-
width mode, but the encoder delivered EVS NB encoded frames.
 IMPACT ON MONITORING
All of this would be of little importance to VoIP monitoring, if the
used bandwidth and bitrate modes had no impact on the user
experience.

The opposite is true. For example, the R-Factor of EVS WB bitrate

modes ranges from 99 (7.2k) to 129 (24.4k), which translates to a
MOSWB range of 3.91 to 4.5. According to ITU-T G.107 this leaves
user satisfaction between “some dissatisfied” to “very satisfied”.

AMR-WB presents an even more extreme example, where satis-

faction (based on wideband reference quality) between the dif-
ferent bitrates ranges from “nearly all users dissatisfied” to “very
satisfied”.

4/9/2019 Duplicate of Max R-Factor by Codec

max wideband ⏵
EVS WB 24.4k 129

EVS WB 32k 129

EVS WB 48k 129

EVS WB 16.4k 126

EVS WB 13.2k 115

EVS WB 9.6k

EVS WB 2.8k 103

107
14
EVS WB 8.0k 103

EVS WB 7.2k 99

EVS NB 24.4k 93

EVS NB 16.4k 90

EVS NB 13.2k 87

EVS NB 9.6k 87

EVS NB 8.0k 83

EVS NB 7.2k 82

EVS NB 2.8k 80
⏴ max narrowband 1/1

This must be seen in light of the fact that very different factors
contribute to codec, codec bandwidth and bitrate selection,
which are not all under full control of a service provider. Quality
can change on a packet by packet basis – put to the extreme, user
satisfaction can change every 20ms.
For modern multi-mode HD voice codecs it is therefore not suf-
ficient to determine the codec based on the result of the SDP
negotiation.

Instead monitoring tools must inspect all RTP packets to reliably

measure the user experience. Yet, most VoIP monitoring tools
focus on the signaling and are unable to cope with the large
amounts of RTP packets.

Real RTP monitoring systems, such as Voipfuture’s Qrystal, analy-

ze the RTP flows and continuously determine the used codecs and
their parameters.

Qrystal uses intelligent timeslicing technology to summarize the

characteristics, such as the codec and its modes, of every five se-
cond time slice. This ensures efficient data storage while providing
a high temporal resolution and accurate estimation of the user
experience.

15
CONCLUSION
For decades narrowband codecs, such as G.711, have been consi-
dered the gold standard of telephony. The current rise of modern
HD codecs is about to change this, which has a number of impli-
cations. For example, HD codecs change the user expectations,
because users quickly get used to new quality standards. This in
turns requires to speed up the transition to wideband codecs.

Technically, HD audio – and especially multi-mode codecs – create

new challenges for service monitoring. Quality can change at any
time; not only because of network impairments, but also because
of codec mode changes. Different codecs and modes may deliver
widely different quality and thus VoIP monitoring systems must be
designed to analyze every RTP packet.

Most VoIP monitoring tools on the market are not capable of de-
tecting the codec based on analysis of the RTP payload. Qrystal
is a notable exception and provides service providers worldwide
with accurate information on the user satisfaction.
Voipfuture is a premium voice quality analytics vendor providing
tools for assessing, aggregating, analyzing, and visualizing voice
quality information. Voipfuture products offer a precise view on
media and control plane to communication service providers,
VoLTE carriers, wholesalers and enterprises.

© 2020 Voipfuture +49 40 688 90 01 0 Wendenstr. 4

All rights reserved, info@voipfuture.com 20097 Hamburg
DOC-ID 144, v1.1 www.voipfuture.com Germany