A Real-Time Transport Protocol (RTP) Extension Header for
Mixer-to-client Audio Level Indication
SIP CommunicatorStrasbourg67000Franceemcho@sip-communicator.orgTelecom ItaliaVia G. Reiss Romoli, 274Turin10148Italyenrico.marocco@telecomitalia.it
This document describes a mechanism for RTP-level mixers in
audio conferences to deliver information about the audio level
of the individual participants. Such audio level indicators are
transported in the same RTP packets as the audio data they
pertain to.
The Framework for Conferencing with the Session Initiation
Protocol (SIP) defined in
RFC 4353
presents an overall architecture for multi-party conferencing.
Among others, the framework borrows from
RTP
and extends the concept of a mixer entity "responsible for
combining the media streams that make up a conference, and
generating one or more output streams that are delivered to
recipients". Every participant would hence receive, in a flat
single stream, media originating from all the others.
Using such centralized mixer-based architectures simplifies
support for conference calls on the client side since they would
hardly differ from one-to-one conversations. However, the
method also introduces a few limitations. The flat nature of
the streams that a mixer would output and send to participants
makes it difficult for users to identify the original source of
what they are hearing.
Mechanisms that allow the mixer to send to participants cues on
current speakers (e.g. the CSRC fields in
RTP) only work for speaking/silent
binary indications. There are, however, a number of use cases
where one would require more detailed information. Possible
examples include the presence of background
chat/noise/music/typing, someone breathing noisily in their
microphone, or other cases where identifying the source of the
disturbance would make it easy to remove it (e.g. by sending a
private IM to the concerned party asking them to mute their
microphone). A more advanced scenario could involve an intense
discussion between multiple participants that the user does not
personally know. Audio level information would help better
recognize the speakers by associating with them complex (but
still human readable) characteristics like loudness and speed
for example.
One way of presenting such information in a user friendly
manner would be for a conferencing client to attach audio level
indicators to the corresponding participant related components
in the user interface as displayed in
.
Implementing a user interface like the above requires analysis
of the media sent from other participants. In a conventional
audio conference this is only possible for the mixer since all
other conference participants are generally receiving a single,
flat audio stream and have therefore no immediate way of
determining individual audio levels.
This document specifies an RTP extension header that allows such
mixers to deliver audio level information to conference
participants by including it directly in the RTP packets
transporting the corresponding audio data.
The key words "MUST", "MUST NOT", "REQUIRED", "SHALL", "SHALL
NOT", "SHOULD", "SHOULD NOT", "RECOMMENDED", "MAY", and
"OPTIONAL" in this document are to be interpreted as described
in RFC 2119.
According to RFC 3550 a mixer
is expected to include in outgoing RTP packets a list of
identifiers (CSRC IDs) indicating the sources that contributed
to the resulting stream. The presence of such CSRC IDs allows an
RTP client to determine, in a binary way, the active speaker(s)
in any given moment. RTCP also provides a basic mechanism to map
the CSRC IDs to user identities through the CNAME field. More
advanced mechanisms, may exist depending on the signaling
protocol used to establish and control a conference. In the case
of the Session Initiation Protocol
for example, the Event Package for
Conference State defines a <src-id> tag which binds
CSRC IDs to media streams and SIP URIs.
This document describes an RTP header extension that allows
mixers to indicate the audio-level of every conference
participant (CSRC) in addition to simply indicating their
on/off status. This new header extension is based on the
"General Mechanism for RTP Header
Extensions".
Each instance of this header contains a list of one-octet
audio levels expressed in -dBov, with values from 0 to 127
representing 0 to -127 dBov(see and
).
Every audio level value pertains to the CSRC identifier
located at the corresponding position in the CSRC list. In other
words, the first value would indicate the audio level of the
conference participant represented by the first CSRC identifier
in that packet and so forth. The number and order of these
values MUST therefore match the number and order of the CSRC
IDs present in the same packet.
When encoding audio level information, a mixer SHOULD include in
a packet information that corresponds to the audio data being
transported in that same packet. It is important that these
values follow the actual stream as closely as possible.
Therefore a mixer SHOULD also calculate the values after the
original contributing stream has undergone possible processing
such as level normalization, and noise reduction for example.
Note that in some cases a mixer may be sending an RTP audio
stream that only contains audio level information and no actual
audio. Updating a (web) interface conference module may be one
reason for this to happen.
It may sometimes happen that a conference involves more than a
single mixer. In such cases each of the mixers MAY choose to
relay the CSRC list and audio-level information they receive
from peer mixers (as long as the total CSRC count remains below
16). Given that the maximum audio level is not precisely defined
by this specification, it is likely that in such situations
average audio levels would be perceptibly different for the
participants located behind the different mixers.
The audio level indicators are delivered to the receivers
in-band using the "General Mechanism for
RTP Header Extensions". The payload of this extension
is an ordered sequence of 8-bit audio level indicators encoded
as per .
The 4-bit len field is the number minus one of data bytes (i.e.
audio level values) transported in this header extension element
following the one-byte header. Therefore, the value zero in this
field indicates that one byte of data follows. A value of 15 is
not allowed by this specification and it MUST NOT be used as the
RTP header can carry a maximum of 15 CSRC IDs. The maximum value
allowed is therefore 14 indicating a following sequence of 15
audio level values.
Note that use of the two-byte header defined in
RFC 5285 follows the same rules
the only change being the length of the ID and len fields.
Audio level indicators are encoded in the same manner as audio
noise level in the RTP Payload Comfort
Noise specification and audio level in the
RTP
Extension Header for Client-to-mixer Audio Level
Notification specification. The magnitude of the audio
level is packed into the least significant bits of one
audio-level byte with the most significant bit unused and
always set to 0 as shown below in
.
The audio level is expressed in -dBov, with values from 0 to 127
representing 0 to -127 dBov. dBov is the level, in decibels,
relative to the overload point of the system, i.e. the
maximum-amplitude signal that can be handled by the system
without clipping. (Note: Representation relative to the overload
point of a system is particularly useful for digital
implementations, since one does not need to know the relative
calibration of the analog circuitry.)
For example, in the case of u-law (audio/pcmu) audio
, the 0 dBov reference would be a
square wave with values +/- 8031. (This translates to 6.18 dBm0,
relative to u-law's dBm0 definition in Table 6 of G.711.)
The URI for declaring the audio level header extension in an SDP
extmap attribute and mapping it to a local extension header
identifier is "urn:ietf:params:rtp-hdrext:csrc-audio-level".
There is no additional setup information needed for this
extension (i.e. no extensionattributes).
An example attribute line in the SDP, for a conference might be:
The above mapping will most often be provided per media stream
(in the media-level section(s) of SDP, i.e., after an "m=" line)
or globally if there is more than one stream containing audio
level indicators in a session.
Presence of the above attribute in the SDP description of a
media stream indicates that some or all RTP packets in that
stream would contain the audio level information RTP extension
header.
Conferencing clients that support audio level indicators and
have no mixing capabilities SHOULD always include the
direction parameter in the "extmap" attribute setting it to
"recvonly". Conference focus entities with mixing
capabilities MAY omit the direction or set it to "sendrecv" in
SDP offers. Such entities SHOULD set it to "sendonly" in SDP
answers to offers with a "recvonly" parameter and to
"sendrecv" when answering other "sendrecv" offers.
The following and show two example offer/answer exchanges
between a conferencing client and a focus, and between two
conference focus entities.
This document defines a means of attributing audio level
to a particular participant in a conference. An attacker may
try to modify the content of RTP packets in a way that would
make audio activity from one participant appear as coming
from another.
Furthermore, the fact that audio level values would not be
protected even in an SRTP session may be of concern in some
cases where the activity of a particular participant in a
conference is confidential.
Both of the above are concerns that stem from the design of
the RTP protocol itself and they would probably also apply
when using CSRC identifiers the way they were specified in
RFC 3550. It is therefore
important that according to the needs of a particular
scenario, implementors and deployers consider use of a lower
level security and authentication mechanism.
This document defines a new extension URI that, if approved,
would need to be added to the RTP Compact Header Extensions
sub-registry of the Real-Time Transport Protocol (RTP)
Parameters registry, according to the following data:
At the time of writing of this document the authors have no
clear view on how and if the following list of issues should
be address here:
Audio levels in video streams. This specification allows
use of audio level values in "silent" audio streams that
don't otherwise carry any payload thus allowing their
delivery within systems where the various focus/mixer
components communicate with each other as conference
participants. The same train of thought may very well
justify audio level transport in video streams.
Roni Even, Ingemar Johansson, Michael Ramalho and several
others provided helpful feedback over the dispatch mailing
list.
SIP Communicator's participation in this specification is
funded by the NLnet Foundation.
During discussions on the subject of audio levels the decision
to transport audio levels in RTP packets, rather than another
protocol was questioned several times which is why the authors
find it worth explaining here. The following subsections
describe alternative mechanisms for delivering audio levels and
the reasons why authors decided not to use them.
RFC 4575 defines a conference
event package for tightly coupled conferences using the
Session Initiation Protocol (SIP) events framework. It
allows for the delivery of various conference related
details such as conference descriptions, participant count
and identity. The document also provides a way of indicating
who the speakers are at any given moment by specifying a
mechanism for mapping conference participants to RTP
SSRC/CSRC identifiers. All these details are dispatched in
an asynchronous manner using the SIP events framework, or,
in other words, through NOTIFY SIP requests following an
initial SUBSCRIBE from a participant.
Contrary to "plain" active speaker infomation, where
significant changes only occur once every several seconds,
audio level in human speech is obviously a very time sensitive
characteristic which would require frequent updates (i.e.
approximately once every 50-100 ms). In order for the update
of the user interface to appear "natural" to the user, audio
level information would probably have to be delivered for
every one or two RTP packets. Using
RFC 4575 or SIP in general for
this would generate traffic on the (often low-bandwidth)
signalling path comparable to, if not exceeding, the media
itself. It may also prove relatively hard for client
developers to synchronize the information they receive from
SIP messages with the one they obtain from the media flows.
It is probably also worth mentioning that the use of
RFC 4575 for such a feature
would make the mechanism incompatible with non-SIP signaling
protocols like, for example,
XMPP and its Jingle extensions.
Similar to using SIP, delivering audio levels through RTCP
would cause bandidth and synchronization issues. Furthermore
the RTP specification explicitly
recommends that the fraction of the session bandwidth added
for RTCP be fixed at 5% which could not be sufficient for
the transport of audio level indicators.
Given the content specific nature of audio levels, it has been
suggested that audio level information be encoded and
transmitted as part of the payload. While this is indeed a
feasible approach, implementing it would require a substantial
effort. In order to implement support for such a feature,
client developers would need to explicitly handle it in all
individual codec modules of their application. Compared to RTP
extensions, the mechanism would therefore represent a
substantial additional effort without offering any meaningful
advantages.
Pulse Code Modulation (PCM) of Voice Frequencies
International Telecommunications Union