The Internet Encyclopedia (Volume 3)

P1: C-172

Kroon WL040/Bidgolio-Vol I WL040-Sample.cls June 20, 2003 13:9 Char Count= 0

316 SPEECH ANDAUDIOCOMPRESSION

Noiseless

decoding

Noiseless

decoding

Inverse

Quantizer

Inverse

Quantizer

Synthesis

filterbank

Synthesis

filterbank

bit

stream audio

output

Figure 13: Block diagram of generic audio decoder.

operation. The resulting bit rate will be variable and will

be signal-dependent. In many implementations an itera-

tive procedure is used to find the optimum quantizer step

sizes that result in coding noise below the masked thresh-

old that will result in the lowest possible bit rate.

The decoder operation performs the operation in re-

verse, without the need for a perceptual model. A generic

block diagram of an audio decoder is shown in Figure 13.

After the coefficients are reconstructed, the signal is trans-

formed back to the time domain and ready for playback.

The block diagrams of Figures 12 and 13 are the princi-

ple for coding a single audio channel. For encoding mul-

tiple channels (Nchannels) once could in principle useN

of these encoder/decoder pairs. However, in practice one

would like to take advantage of the possible correlations

that exist between the various channels. Also, for trans-

parent coding, one would take into account that masking

levels will differ for signals that are spatial in nature. Some

distortions that are inaudible in each individual channel

will become audible when listening to its multichannel

version (e.g., in stereo).

Most state-of-the-art audio coders will produce a vari-

able bit rate. In most communication scenarios a fixed bit

rate is more desirable. This is accomplished by a buffering

scheme that interacts with the coder quantization deci-

sions. Designing buffering schemes that minimize buffer

size (and its corresponding delay) and minimize impact

on audio quality turns out to be a challenge, and various

solutions exist, each with advantages and disadvantages.

Audio Coding Standards

Two types of standards exist. The first type is based on

sanctioning by a standard organization (e.g., ISO and

its MPEG standards). The second type is based on a

proliferated proprietaryde factostandard (Windows Me-

dia Player, RealPlayer, Dolby AC-3, etc). In both cases

proper licensing is needed to use these standards in com-

mercial applications. The MPEG coding standards are

widely used for audio and video. The well-known MP3

standard is actually MPEG-1, Layer 3. Table 4 summa-

rizes the MPEG standards.

The MPEG standards are defined in a different way

than most ITU-T speech coding standards. They consist of

a normative part defining the bit stream syntax and the de-

coder description (sometimes accompanied by reference

C code). As a result anyone should be able to implement

proper decoders. The encoder is described in the informa-

tive part and just describes the algorithmic concepts and

operating modes. However, to build good encoders it is

necessary to understand the algorithms, and the standard

document will not provide details on how to build a good

encoder. As a result standard compliant encoders can be

quite different in performance.

Besides coders based on the traditional coding

paradigm described above, other paradigms exist as well.

These alternative paradigms play a role in very-low-rate

coding. Two paradigms that have become part of the

MPEG-4 standards are structured audio and parametric

audio coding. Structured audio consists of a set of tools

that can be used to generate music in a way similar to a

music synthesizer. It also contains a structured way of de-

scribing musical scores and their subsequent conversion

to sound using synthesizers. A structured audio bit stream

describes how to build the synthesizers and provides the

musical score and information on how to play this on the

synthesizer. The resulting description can be very com-

pact and low-bit-rate (only several kb/s). The resulting au-

dio signal can be of very high quality, and because many

modern music productions use synthesizers extensively, it

can be very close to some originals. Parametric audio cod-

ing uses ideas similar to those used in speech compression

by modeling some of the production mechanisms of the

music sounds. Although it will work well for some sounds,

it is difficult to make this technique work in a consistent

way for a wide variety of input signals.

Table 4Overview of Various MPEG Audio Standards

Standard Year Rates for transparency Channels Comments

MPEG-1 Audio, 1992 384 kb/s Mono, stereo 48-, 44.1-, 32-kHz

Layer I stereo sampling rates

MPEG-1, Layer II 1992 192—256 kb/s Mono, stereo 48-, 44.1-, 32-kHz

stereo sampling rates

MPEG-1, Layer III 1992 128—160 kb/s Mono, stereo 48-, 44.1-, 32-kHz

stereo sampling rates

MPEG-2, Layers I, 1994 See MPEG-1 Mono, stereo, and MPEG-1 + enhancements,

II, III rates backward compatible supports lower sampling

5.1 multichannel rates

MPEG-2, AAC 1997 96—128 Mono, stereo, multichannel Not compatible with

up to 48 channels MPEG1, 2, supports

96-kHz sampling

MPEG-4 Version 1 1998 9—128 Mono, stereo, multichannel Supports various coding tools

MPEG-4 Version 2 1999 9—128 Mono, stereo, multichannel Support error robustness tools