HE-AAC Decoder

1
HE-AAC Decoder

• Detail and Standard

Speaker ???
2
Outline

• Introduction
• Concept of SBR
• SBR Block Diagram
• First Step T/F Grid
• Envelope Decoding
• QMF
• HF Generation
• HF Adjustment
• Low power SBR
• Conclusion

3
Introduction

• Bandwidth Extension, also called SBR, is a
  synthesis algorithm to rebuild the lost
  bandwidth.
• The basic idea is
• There exists a strong correlation between the
  high band characteristics of a signal with the
  low band characteristics of the same signal
• Several bandwidth extension algorithms have been
  proposed for Speech Coding, however they are not
  suitable for Audio coding
• The first well-known audio codec with bandwidth
  extension is MP3Pro, which is the combination of
  MP3 and bandwidth extension.

4
Introduction (cont.)

• HE-AAC
• High efficiency AAC
• About 1-3 kbps/ch side information
• Coding efficiency improves at least 30
• Developed by Coding Technologies
• SBR is a post process
• Either MPEG-4 or MPEG-2 AAC can cooperate with
  SBR

5
Introduction (cont.)

• Low power SBR
• A simplified SBR
• Use real-valued filterbank to replace complex QMF
• Add Aliasing Reducer to reduce aliasing
• The SBR algorithm in MPEG-4 is still a simple
  method, and, thus, it may not be suitable to all
  case.
• MPEG is still trying to propose new SBR technique
  to improve the robust of SBR.
• Patch The action that copies/reproduces data
  from low freq. to high freq.

6
Concept of SBR

• Reconstruct high frequency data
• Use low frequency to reconstruct high frequency
• Copy low frequency to high frequency
• Adjust its envelope

Adjust envelope
7
Concept of SBR (cont.)

• Complex-Exponential Modulated Filter Banks
• Adding an imaginary part to the real-valued
  filter bank
• Main alias terms no longer exist
• Enhancements
• Transposition can only solve stationary
  components
• Two enhancement model
• Strong tonal components in high frequency
• Tone-to-noise ratio differs between high/low
  frequency

8
Idea of Low Power SBR

• Reason why aliasing occurs
• SBR independently modify subband gains.
• Source of aliasing
• Envelope adjustment
• Tone injection
• ADD Aliasing Reducer block
• Group bands and give
• them same gain
• Subtract the approximation
• of aliasing term from
• tone injection

Gain modified
9
SBR Block Diagram

• SBR Decoder

10
SBR Block Diagram (cont.)

• Decoding step

Audio signal
QMF
T/F grid
SBR data
Envelope Decoding
HF Gen.
HF Adj.
11
First Step T/F Grid

• The first step of SBR decoding is to setup
• Start/Stop freq. of SBR frame or patch borders.
• the T/F grid for this SBR frame.
• The grid is
• Unit of gain modification
• Can view as grouping
• Freq. tables recode the start pointer of each
  bands/group

12
T/F Grid Freq. Part
1
QMF
32
Grid
13
T/F Grid Freq. Part (cont.)

• Three basic borders
• Start of , also the end of data
  been used to patch (see later)
• Start of SBR data
• End of SBR data
• The fundamental freq. table.

14
T/F Grid Freq. Part (cont.)

• Table derivations and their relations

15
T/F Grid Freq. Part (cont.)

• Calculate
• According to and
• The scale may be linear or 8, 10, and 12
  octaves/band
• Equal distance of each band by their scale in
  each section
• Number of bands are set to even by group some
  bands. ( Why? )

16
T/F Grid Freq. Part (cont.)

• Calculate

No grouping
Linear
Group 2 channels
1 section
Octave Based
No alteration
2 section
High bands alteration (lower resolution)
17
T/F Grid Freq. Part (cont.)

• Calculate
• Calculate
• Subset of , similar to downsampling
  by a factor of 2
• Calculate
• Subset of
• Number of bands is controlled by encoder

18
T/F Grid Freq. Part (cont.)

• Calculate
• Used for limiter of gain
• By Merging Patchborders with

19
T/F Grid Time Part

• 4 kind of mode FIXFIX, FIXVAR, VARFIX, and
  VARVAR
• The mode VAR needs Encoder to transmit the
  relative starting borders and lengths.

Starting border
End border
20
Envelope Decoding

• Decoding Step

Lossless Decoder
Dequant
Stereo decoding
21
Envelope Decoding (cont.)

• Lossless Decoder
• Delta Coding
• Encode the differencedelta
• Use Huffman coding to encode the difference
• Two kinds of direction
• Time
• Frequency
• For time direction of signal envelopes
• across SBR envelope
• Should care about the transition of time slots
  with different freq. resolution.

22
Envelope Decoding (cont.)

• Dequant
• Signal envelope
• a 1 or 2, according to header
• Noise envelope
• Stereo decoding
• Not coupling gt as single channel
• Coupling
• Two sets of Envelopes
• First one is Envelope average
• Second one is ratio of left to right channels

23
QMF

• Analysis filterbank
• 320 overlapped samples for each analysis
  filtering
• Step

Data 320 samples
Windowing
Sum up into 64 samples
Transform
24
QMF (cont.)
642
25
QMF (cont.)

• Synthesis filterbank

Inverse transform
128
128
128

Re-arrange
Windowing
Sum up into 64 samples
26
HF Generation

• Two functions
• Inverse filtering
• Patch
• Inverse filtering
• 2nd order filtering for each subband
• Encoder can set parameters to control if the
  relation to previous 2 slots is large
• Patch

Patch 2
Patch 1
Patch 2
Patch 1
27
HF Adjustment

• Functions
• Add sinusoids
• Adjust gain to meet envelope from envelope
  decoding
• Add all components
• Sinusoids adding
• A vector recode if sets of T/F grids contain
  missing sinusoids
• Add a sinusoid in the middle of Freq grid

28
HF Adjustment (cont.)

• Adjust gain
• Estimation of current envelope (patched data
  envelope).
• Modify envelope by additional sinusoids and noise
  components
• Calculate gain

29
HF Adjustment (cont.)

• Add all components
• Multiply gain to
• Use table to find noise sample values and apply
  gain to them
• apply gain to sinusoids. (sinusoid Real part
  1,0,-1,0,1,0,)

30
Low power SBR

• Low power SBR Decoder

31
Low power SBR (cont.)

• Aliasing detection
• Make use of reflection coefficients
• Two conditions to decide aliasing degree
• degree 1 if
• degree 1 rek(k-1)2

32
Low power SBR (cont.)

• Aliasing reduction
• Group the aliasing term together
• Each group will not contain
• Degree 0, that is to say, no aliasing
• A sinusoid add in this subband
• Modify gain to let each component in the same
  group has the same gain
• Calculate aliasing term from additional sinusoids
• Subtract this aliasing term

33
Conclusion

• SBR needs just 1-3 kbps/ch side information to
  synthesis high frequency data
• Dramatically reduces data rate
• Coding efficiency improves at least 30
• Avoiding aliasing is important to generate high
  frequency data.
• The model to generate high frequency data is
  still an easy one
• T/F grids adaptively match the suitable
  resolution of input data envelope.
• More SBR techniques may be proposed in the future.

34
Reference

• Per Ekstrand, Bandwidth Extension of Audio
  Signals by Spectral Band Replication, IEEE
  Benelux Workshop on Model based Processing and
  Coding of Audio (MPCA-2002), Leuven Belgium, Nov
  15, 2002
• Martin Wolters1, Kristofer Kjorling2, Daniel
  Homm1, Heik Purnhagen2, A closer look into
  MPEG-4 High Efficiency AAC, 115th AES
  Convention, Munich, October 2003.
• Chong Kok Seng, etc., Low Power Spectral Band
  Replication Technology for MPEG-4 Audio
  Standard, ICICS-PCM, Dec., 2003
• Martin Dietz, MPEG-4 Extension 1 Bandwidth
  Enhancement, 114th AES Convention, March, 2003.
• ISO/IEC JTC/SC29/WG11/N7128, MPEG-4 Audio Third
  Edition