INTERNATIONAL ORGANISATION FOR STANDARDISATION

ORGANISATION INTERNATIONALE DE NORMALISATION

ISO/IEC JTC1/SC29/WG11
CODING OF MOVING PICTURES AND AUDIO
ISO/IEC JTC1/SC29
WG11 N15073
February 2015, Geneva, Switzerland

Title Technical Note: AAC-ELD v2 IMPLEMENTATION GUIDE

Source Communication

AhG on Responding to Industry Needs on Adoption of MPEG Audio

The purpose of this document is to describe principles of explicit AAC-ELD v2 signaling and
filterbank set-up. The intention is to support those implementing AAC-ELD codecs with an
overview of the relevant standards as well as some implementation tips and hints. It is in no
way intended to replace intensive study of MPEG reference software and standard documents.

1. Introduction
The AAC-ELD family includes Low Delay AAC (AAC-LD), Enhanced Low Delay AAC
(AAC-ELD) and AAC-ELD v2. The order of the three family members represents an
evolution in AAC low delay audio codec technology. These codecs share the same core coder
while each descendant adds new coding tools (Figure 1). Decoders of the AAC-ELD codec
family can be expected to be fully backwards compatible. The codecs can handle mono,
stereo and multichannel signals with latencies as low as 15ms at 48 kHz and support for a
wide range of bit rates.

Figure 1: AAC-ELD family. AAC-LD and AAC-ELD share all major tools, only
minor differences of the bit stream syntax and filterbank separate AAC-LD from
AAC-ELD core. AAC-ELD v2 is based on AAC-ELD and is extended by the Low
Delay MPEG Surround tool. (Image source: Fraunhofer IIS)
AAC-ELD v2 is the latest member of the AAC-ELD codec family. It is based on AAC-ELD
and is extended with the “Low Delay MPEG Surround” (LD-MPS) tool, which provides
improved audio quality for stereo and multichannel at low bit rates.

Instead of discrete channel coding, the AAC-ELD v2 encoder extracts spatial parameters and
encodes a downmix stream (Figure 2). The AAC-ELD core decoder reconstructs the
downmix signal while the LD-MPS decoder module recreates the spatial image by applying
the LD-MPS parameters to the downmix signal (see[1][2]).

Figure 2: Block diagram of AAC-ELD v2 (Image source: Fraunhofer IIS)

2. Background
2.1 List of Abbreviations
AAC Advanced Audio Coding
AAC-ELD Enhanced Low Delay Advanced Audio Coding
AAC-LD Low Delay Advanced Audio Coding
AAC-(E)LD AAC-LD and AAC-ELD
AAC-ELD v2 Enhanced Low Delay Advanced Audio Coding Version 2
AOT Audio Object Type
ASC Audio Specific Configuration
CLDFB Complex Low Delay Filterbank
DEMUX De-Multiplexer
iMDCT Inverse Modified Discrete Cosine Transform
LD-MDCT Low Delay Modified Discrete Cosine Transform
LD-MPS Low Delay MPEG Surround
LD-QMF Low Delay Quadrature Mirror Filter
LD-SBR Low Delay Spectral Band Replication
MDCT Modified Discrete Cosine Transform
MPEG Motion Picture Experts Group
MUX Multiplexer
PCM Pulse Code Modulation
PNS Perceptual Noise Substitution
QMF Quadrature Mirror Filter
SBR Spectral Band Replication
SIP Session Initialization Protocol
TNS Temporal Noise Shaping

2.2 Reference Documents

As the AAC-ELD v2 coding scheme is based on its predecessors, AAC-LD and AAC-ELD,
there is no single document in the MPEG standard describing all AAC-ELD v2 coding tools
in one place. For AAC-ELD v2, the relevant MPEG standard documents are:

 ISO/IEC 14496-3:2009. This MPEG-4 standard defines the basic AAC-LD and
AAC-ELD coding schemes (AOTs 23 and 39) and its signaling via the
AudioSpecificConfig() and the ELDSpecificConfig() [3].
 ISO/IEC 14496-3:2009/Amd 2:2010. This amendment to the 14496-3 standard
contains signaling of LD-MPS (AOT 44) via the LDSpatialSpecificConfig()
[4].
 ISO/IEC 14496-3:2009/Amd 3:2012. This amendment to the 14496-3 standard
defines the Low Delay AAC v2 Profile [7].
 ISO/IEC 23003-1:2007. Basic tools and bit stream elements are defined in the MPEG
Surround standard [6].
 ISO/IEC 23003-2:2010. This standard is part of MPEG-D and describes the syntax of
the LDSpatialSpecificConfig() and the LD-MPS payload [5].
2.3 Terminology
Coding Tools. Each MPEG codec is a composition of one or more coding tools/modules. The
important tools are Low Delay Window, Filterbank, Quantization&Coding – AAC, MS
Stereo, SBR and others. For a complete list, please see table 1.1 “Audio Object Type
definition based on Tools/Modules” in [3].

Audio Object Type, referred to as AOT. An Audio Object Type is a self-contained codec
which consists of a defined set of coding tools. For instance, AOT “ER AAC-ELD” has the
ID 39 which includes the tools Low Delay Window, Filterbank, TNS, Intensity Stereo, MS
Stereo, PNS, Quantization&Coding, Low Delay SBR and others. Table 1 describes AOTs 23,
39 and 44. For a complete overview of all AOTs, please see table 1.1 “Audio Object Type
definition based on Tools/Modules” in [3] and [7].

AOT Description
23 Low Delay AAC (AAC-LD)
39 Enhanced Low Delay AAC (AAC-ELD)
Enhancement of AAC-LD, with an delay optimized filterbank and the Low Delay
Spectral Band Replication tool (LD-SBR) for low bit rate encoding
44 Low Delay MPEG Surround (LD MPEG Surround) for parametric coding of
stereo and multichannel signals.
Table 1: Description of AOT 23, 39 and 44

 Audio Profiles. An Audio Profile provides an application-oriented set of one or more

AOTs and may also restrict tools. Audio Profiles are signaled during the codec
negotiation and capability exchange amidst session setup (e.g. via Session
Initialization Protocol (SIP)). Table 2 shows relevant Audio profiles for the AAC-ELD
family. For a complete overview of all MPEG Audio Profiles, please see section
1.5.2.1 “Profiles” in [3] and [7].

MPEG Audio Profile Associated AOTs

Low Delay 23 (AAC-LD) and others
Low Delay AAC 23 (AAC-LD)
Low Delay AAC v2 23 (AAC-LD)
39 (AAC-ELD)
44 (LD MPEG Surround)
Table 2: MPEG Audio Profiles relevant for the AAC-ELD family

 Audio Profile Levels. These levels allow signaling of restrictions within an Audio
Profile. The Low Delay AAC v2 profile consists of 4 levels which are differentiated in
the number of output channels and in the low delay MPEG surround channel
configuration (Table 3). For a complete overview of all Audio Profile Levels, please
see section 1.5.2.3 “Levels within the profiles” in [3] and [7].

Level AOT of core Max. no. Max. sampling MPEG

coder output rate [kHz] Surround
channels
1 23, 39 1.0 48 n/a
2 23, 39 2.0 48 LD-MPS 2-1-2
3 23, 39 5.1 48 LD-MPS 2-1-2
 4 23, 39 5.1 48 LD-MPS 5-x-5
Table 3: Levels of the Low Delay AAC v2 profile

 Payload. The payload contains the coded audio data including tools like SBR or LD-
MPS. Different AOTs use different bit stream payload syntax.

3. Signaling of AAC-ELD v2
3.1 Signaling MPEG-4 Codecs
A decoder requires configuration information to set up its internal configuration for proper
parsing and decoding of the bit stream. The configuration information is packed into the
Audio Specific Configuration (ASC) and sent to the decoder before the audio data bit stream.
A complete description of the ASC syntax can be found in section 1.6.2.1
“AudioSpecificConfig” in [3].

3.2 The Audio Specific Configuration for AAC-LD and AAC-ELD

In case of AAC-ELD (AOT 39) the ASC includes the ELDSpecificConfig() element.
The exact syntax can be found in table 4.180 “Syntax of ELDSpecificConfig ()” in [3].

In case of AAC-LD (AOT 23) the ASC includes the GASpecificConfig(). The exact
syntax can be found in section 4.4.1 “Decoder configuration (GASpecificConfig)” in [3].

3.3 The Audio Specific Configuration for LD-MPS

The configuration of the LD-MPS tool is stored in the LDSpatialSpecificConfig()
as described in section B.2.1 “Payloads for LD MPS” in [5]. It is embedded into the ASC but,
depending on the underlying core codec (AAC-ELD or AAC-LD), the
LDSpatialSpecific-Config() has to be accessed in two different ways that are
explained in the following sections 3.3.1 and 3.3.2.

3.3.1 Signaling AAC-ELD v2 (AAC-ELD and LD-MPS )

When using AAC-ELD (AOT 39) as core codec, the LDSpatialSpecificConfig() is
embedded in the ELDSpecificConfig() (see table 4.180 “Syntax of ELDSpecificConfig
()” in [4]).
Figure 3: Signaling AAC-ELD v2 (Image source: Fraunhofer IIS)

3.3.2 Signaling AAC-LD and LD-MPS

When using AAC-LD (AOT 23) as core codec, the LDSpatialSpecificConfig() is
embedded in the AudioSpecificConfig() (see table 1.15 “Syntax of
AudioSpecificConfig()” in [4]).

Figure 4 shows the hierarchical structure of the ASC when signaling AAC-LD and LD-MPS.

Figure 4: Signaling AAC-LD and LD-MPS (Image source: Fraunhofer IIS)

3.4 The LD-MPS Payload within the AAC Bit Stream

Every AAC frame consists of the AAC-(E)LD core data and, if applicable, also the SBR data
and the extension payload. The SBR data is stored directly after the core data while the LD-
MPS payload is embedded in the extension_payload() and classified as extension type
EXT_LDSAC_DATA (see table 4.57 “Syntax of extension_payload()” in [4]).

To ensure forward compatibility the extension type EXT_DATA_LENGTH has been

introduced together with LD-MPS. This extension type defines a payload that contains the
explicit length of the subsequent payload container and thus enables skipping of
unknown/future extension payloads.

Figure 5 shows the structure of an AAC-ELD bit stream with embedded LD-MPS payload.

Figure 5: The LD-MPS payload within the AAC-ELD bit stream (Image source:
Fraunhofer IIS)
Note: For AAC-LD the payload structure is identical, although it contains no SBR data.

4. AAC -ELD v2 Filterbank dependencies

As shown in Figure 2, an AAC-ELD v2 codec includes the AAC-ELD core codec, the LD-
SBR and the LD-MPS tool. According to the MPEG standard, each of these components uses
another type of filterbank as listed in Table 4.

Tool Used type of filterbank

AAC-ELD core codec LD-MDCT
LD-SBR CLDFB
LD-MPS LD-QMF
Table 4: The different types of filterbanks used in the AAC-ELD v2 codec

To save workload it is possible to employ the LD-QMF also for the LD-SBR tool and to omit
the CLDFB. In this section there is a full explanation of the different variants of sequential
arrangements of AAC-ELD v2 filterbanks for the encoder and decoder.

4.1 Filterbank Sequence of the AAC-ELD v2 Encoder

4.1.1 Extending a Legacy Encoder: Each Tool Has Its Own Filterbank
This implementation option shows how a legacy AAC-ELD encoder can be extended by the
LD-MPS tool without any modification to the legacy part. As Figure 6 shows, the LD-MPS
and LD-SBR tools do not exchange any frequency domain data. Two separate filterbanks -
LD QMF analysis in the LD-MPS tool and CLDFB analysis in the LD-SBR tool - must be
processed.

Fi
gure 6: Extending a legacy AAC-ELD encoder by the LD-MPS tool (Image
source: Fraunhofer IIS)

The LD-QMF causes a higher delay than the CLDFB. When LD-SBR operates in Dual Rate
mode the LD-SBR path has to be delayed by 96 samples to align the LD-SBR to the LD-MPS
tool. If the SBR tool runs in Down Sampled SBR mode 1, the LD-SBR path must be delayed
by 48 samples.

1
An explanation of Down Sampled LD-SBR can be found in [1] and in 4.6.18.2.1.3 “Down Sampled SBR” in
[3].
4.1.2 Workload Efficient implementation: Sharing LD-QMF Data between tools
As depicted in Figure 7 the LD-MPS tool directly shares its LD-QMF filterbank output data
with the LD-SBR encoding block. Therefore, it is possible to omit the LD-SBR CLDFB
analysis block that is usually part of the LD-SBR encoder. In this way workload of the
encoder can be reduced compared to the implementation explained in section 4.1.1.

Figure 7: Workload efficient implementation achieved by sharing the output of

the LD-QMF with the LD-SBR encoder (Image source: Fraunhofer IIS)

Here the LD-SBR and LD-MPS signal paths share the same LD-QMF filterbank, and
therefore no delay alignment between the LD-SBR and LD-MPS tool is required.

4.2 Filterbank Sequence of the AAC-ELD v2 Decoder

When adding LD-MPS to an AAC-ELD v1 decoder a LD-QMF filterbank has to be
implemented. This filterbank is used for LD-MPS processing and also for LD-SBR processing
whenever LD-MPS has been activated on encoder side. The CLDFB is exclusively active for
plain LD-SBR decoding, i.e. when LD-MPS is inactive. The structure of the AAC-ELD v2
decoder is depicted in Figure 8, where we have to differentiate between the following cases:

 LD-MPS off / LD-SBR off. If no tool is signaled as active, the iMDCT of the AAC
core directly outputs the decoded signal. This case is not explicitly shown in Figure 8.
 LD-MPS off / LD-SBR on. If LD-SBR is signaled as active and if ELDEXT_LDSAC
is not present in the ELDSpecificConfig() (see also section 3.3.1 and 4.6.20.4
“Decoding of ELDSpecificConfig” in [4]), the CLDFB is used for LD-SBR analysis
and synthesis.
 LD-MPS on / LD-SBR off. If LD-SBR is not present in the bit stream but
ELDEXT_LDSAC exists, the LD-QMF is used for LD-MPS analysis and synthesis.
 LD-MPS on / LD-SBR on. If both tools are signaled as active in the bit stream, the
output of the LD-QMF analysis filterbank is directly used by the LD-SBR decoder
followed by the LD-MPS decoder and LD-QMF synthesis filterbank. The CLDFB is
not active in this case.

The necessary filterbank depending on the active tool is also shown in Table 5.

LD-MPS off LD-MPS on

LD-SBR off N/A LD-QMF
LD-SBR on CLDFB LD-QMF
Table 5: Used filterbank in an AAC-ELD v2 decoder depending on the active
tools

Figure 8: Filterbank sequence of the AAC-ELD v2 decoder (Image source:

Fraunhofer IIS)

4.3 Forward Compatibility of AAC-ELD v1 Decoders

AAC-ELD v1 is forward compatible to AAC-ELD v2, i.e. when receiving an AAC-ELD v2
bit stream an AAC-ELD v1 decoder does not process the LD-MPS payload but outputs the
decoded mono AAC core and LD-SBR parts of the signal. As a result, the LD-SBR signal
lags slightly behind the AAC core signal 2 due to the AAC-ELD v2 encoder LD-SBR delay
alignment of 48 or 96 samples (see section 4.1.1). The alignment has been done under
the assumption that the decoder utilizes the LD-QMF for LD-SBR processing. Hence, the
time domain input of the CLDFB analysis should be delayed by 96 samples in Dual Rate SBR
mode or 48 samples in Down Sampled SBR mode to compensate for the AAC-ELD v2
encoder delay alignment.

It is recommended to update AAC-ELD v1 decoders as follows (see Figure 9):

 Adapt the ASC parser of the legacy decoder to detect the eldExtType
ELDEXT_LDSAC in the ELDSpecificConfig(). If ELDEXT_LDSAC exists, set
the flag useLDQMFTimeAligment3 to 1.
 Implement a delay shift of the time domain input buffer for the LD-SBR CLDFB
analysis by the appropriate number of samples (dual rate SBR: 96 samples; Down
Sampled SBR: 48 samples). This alignment has to be triggered by the
useLDQMFTimeAligment flag. For details see 4.6.20.4 “Decoding of
ELDSpecificConfig” in [4].
 It is also recommended to support the extension payload EXT_DATA_LENGTH that
enables skipping of unknown/future extension payloads (see section 3.4).

2
Listening tests during the standardization process showed that this delay shift has almost no impact on the
audio quality.
3
In the spelling of “useLDQMFTimeAligment” the “n” lacks by intention. You will find the same typo in the
standard.
Figure 9: AAC-ELD v1 decoder dealing with AAC-ELD v2 bit streams (Image
source: Fraunhofer IIS)

5. Conclusion
When implementing AAC-ELD v2 codec based on the AAC-ELD v1 codec the following
modifications have to be conducted:

 ASC writer / parser. The configuration of the LD-MPS tool has to be stored in the
LDSpatialSpecificConfig() that is embedded in the
ELDSpecificConfig() for AOT 39, respectively in the ASC for AOT 23.
 LD-MPS payload. The LD-MPS data has to be stored in an extension payload
container of the type EXT_LDSAC_DATA. Additionally, the payload
EXT_DATA_LENGTH should be added to facilitate forward compatibility to possibly
“unknown” extension payloads.
 Filterbank sequence of the AAC-ELD v2 encoder. Option one: Every tool uses its
own filterbank. In this case the LD-SBR signal part has to be delayed to be in sync
with the LD-MPS tool. Option two: The LD-MPS tool shares its LD-QMF data with
the SBR tool. This is the recommended implementation as the workload is reduced
compared to option one.
 Filterbank sequence of the AAC-ELD v2 decoder. Whenever LD-MPS is active the
LD-QMF has to be used for processing LD-MPS as well as for LD-SBR. The CLDFB
is exclusively used for plain LD-SBR decoding when LD-MPS is inactive.
 AAC-ELD v1 decoders. Legacy AAC-ELD v1 decoders are forward compatible to
AAC-ELD v2 bit streams but cause a slight delay shift between the LD-MPS and the
LD-SBR signal parts. For compensation it is recommended to update the ASC parser
to detect the presence of LD-MPS and to introduce the appropriate delay to the LD-
SBR signal part.
Further references
[1] Mpeg-audio.org White Paper: The AAC-ELD family for High Quality
Communication Services. http://www.mpeg-audio.org/docs/w14936_(AAC-ELD-
Family_WhitePaper).pdf
[2] Maria Luis Valero, Andreas Hoelzer, Markus Schnell , Johannes Hilpert, Manfred
Lutzky, Jonas Engdegard, Heiko Purnhagen, Per Ekstrand, Kristofer Kjoerling. A new
parametric Stereo and Multi Channel Extension for MPEG4 Enhanced Low Delay
AAC (AAC-ELD v2). In 128th AES convention, London, UK, May 2010.
[3] ISO/IEC 14496-3:2009. Information technology -- Coding of audio-visual objects –
Part 3: Audio.
[4] ISO/IEC 14496-3:2009/Amd 2:2010. ALS simple profile and transport of SAOC.
[5] ISO/IEC 23003-2:2010. Information technology -- MPEG audio technologies - Part 2:
Spatial Audio Object Coding (SAOC).
[6] ISO/IEC 23003-1:2007. Information technology -- MPEG audio technologies -- Part
1: MPEG Surround.
[7] ISO/IEC 14496-3:2009/Amd 3:2012. Transport of unified speech and audio coding
(USAC)