Chapter 5 Summary: Review of MPEG-4 General Audio Coding

Acronyms:
LPC: Linear Predition Coding
PCM: Pulse-Code Modulation
DPCM: Diff PCM
ADPCM: Adaptive DPCM
CELP: Codebook Excited Linear Prediction

Two types of coding:

- General Audio Coding: general audio coding
- Synthetic Audio Coding: synthetic creation and parametric description of an audio signal

A Codecs

source coding
generally lossy, not very good for (music).

Perceptual coding
Removes parts of the signal that are masked by other signals or simply below hearing threshold.

LPC is described, whose main modules are the difference calculation, the quantizer and the predictor. the error between the prediction and the real signal is what's transmitted

A sub-band codec uses a filter bank which splits different frequencies into bands which are then quantized using fewer bits since sub-bands have smaller steps than the original

Perceptual coding makes use of a perceptual model. It attempts to partition frequency (and time) to match the capacity of the human hearing. The rate (loop?) used into such codecs are used to optimize rate versus quality.

2. MPEG-2 AAC Advanced Audio Coding

Directly incorporated into MPEG4 General Audio Standard.
Supports:
- multiple audio channels
- subwoofer channels
- embedded data channels
Excellent quality at rates of 64kbps/channel and less

It uses
- high resolution filter bank
- backward adaptive prediction
- joint channel coding
- Huffman coding
- sampling rates -> from 8 to 96 khz
- wide range of bit rates

The block diagram from page 136 introduces a few elements

Noise Shaper: work with the Filter Bank by modifying it's characteristics
M/S: increases the coding gain by removing inter channel redundancy, preserve stereo/multichannel signal from noise.

Rate-distortion control: adjust the scale factor to adjust the overall bit rate

Analysis-synthesis filter bank

Stationary signal:
- high frequency rezolution (better prediction gain)
- critical sampling (***# of ts = # of fc***) minimize the # of fc to quantize
- *overlap-add* synthesis (smooths the quantization)
- perfect reconstruction

Transient signal:
- high time resolution
- short block mode (describe)

DRAWING

Perceptual model
??

Quantization and Noiseless Coding

AAC is a variable-rate codec, so the rate distortion module adjusts step size and # of quantization level to achieve *constant* rate

 non-linear quantization (Huffman coded - 11 different codebooks, can be used for the same spectrum)
 rate-distortion
 noiseless coding

MPEG4 Addition to AAC

new tools within the codec:
- PNS (Perceptual Noise Subsystem)
Based on similarity between noise like signals. The noiselike signal is not quantized. Instead a flag is set and the total power is tranmitted for each band.

- LTP (Long Term Prediction)
Select between transmitting the residual signal or the original signal for a band.

- TwinVQ (Transform Domain Weighted Interleaved Vector Quantization)
for low bitrates, at or less than 16 kbits/sec

MPEG4 Scalabel General Audio Coding

This allows to encode independently of any bit rate requirement. Bit streams generated consist of several partial bitstream encoded separately.
Layered coding where each layer uses the difference between the original signal and the decoded signal of the previous layer. Typically 3-4 layers.

MPEG4 Scalable Audio Coder

Having a CELP core coder and ACC enhancement layer is the best configuration for speech quality.
Target application:
- merging of all audio service
- removes transcoding needs
- heterogeneous receivers

General Characteristics:

- 2kbits/sec to 100kbits/sec per channel
- layered coding
- MPEG-4 speech coder or MPEG-4 General Audio'coder
- multiple sampling rate
- multiple bitrates options