The Internet Encyclopedia (Volume 3)

P1: C-172

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

AUDIOCODINGTECHNIQUES 315

of this process. The auditory system has been shown to

have masking properties. Masking describes the process

in which one signal becomes inaudible in the presence

of another signal. In other words, under certain circum-

stances it is possible to make quantization noise inaudi-

ble while the decoded audio signal is present. Masking

can happen both in time and in frequency. Understand-

ing the principles of masking has taken many decades of

research using both physiological and physical measure-

ments. Masking data are obtained through psychoacous-

tical studies, in which subjects are exposed to test signals

and asked about their ability to detect changes (increase

in frequency, audibility, etc). Most of the understanding

of masking is based on simple tones and noise. Because

complex signals can be viewed as composites of time-

varying tones, a common approach has been to derive the

masked threshold by analyzing the signal tone by tone in

specific frequency bands, related to the frequency bands

used by the human auditory system to analyze sounds.

These bands, called critical bands, are spaced nonuni-

formly with increasing bandwidth for higher frequencies.

In each critical band, the signal and its corresponding

masking function are calculated, and the masked thresh-

old is derived as a superposition over the complete fre-

quency band. It should be noted that the actual procedure

is much more complicated, taking into account several

interactions and characteristics of the signals (e.g., if the

signal is noiselike or tonelike). Figure 11 gives an example

of the power spectrum of a signal and its corresponding

masked threshold. In this figure, as long as the quantiza-

tion noise remains below the solid line, it will be inaudible.

In general the model that is used to derive the masked

thresholds is referred to as thepsychoacousticorpercep-

tual model. Building a good perceptual model and match-

ing it properly to a given coder structure is a very complex

task. It also should be noted that for a given coder struc-

ture (or coder standard such as MPEG 1, Layer 3), it is

possible to improve the perceptual model while still being

compliant to the standard. This also explains the quality

differences in various encoders that all support the same

standard.

To achieve redundancy removal, and at the same time

take advantage of the frequency-domain masking proper-

ties, it is beneficial to perform a spectral decomposition

60

Power (dB)

-40

0Hz 16kHz

Frequency

Figure 11: Example of signal frequency power spectrum and

its corresponding masked threshold (solid stepped line).

of the signal by means of a filter bank or transform. Most

modern audio coders are based on some form of lapped

transform, which not only provides computational effi-

ciency, but also allows perfect reconstruction. In other

words the transform and its inverse will produce a de-

layed version of the original time signal. The transform

sizes and overlaps will be chosen in such a way that the

signal is critically sampled (i.e., the number of frequency

components is the same as the number of time samples).

The size of the transform will determine the spectral res-

olution. A larger transform will provide better spectral

resolution at the expense of decreased time resolution. A

common solution is to make the transform size adapt to

the signal. For stationary signals, a large transform size

is chosen, whereas for nonstationary signals or onset’s a

smaller transform size is chosen. Typically sizes vary from

256 to 2,048 samples for sampling rates in the range 20

to 40 kHz. This process is called window switching, and

care is taken to make sure that the whole process is invert-

ible. A commonly used transform is the modified discrete

cosine transform (MDCT). It uses a transform of length

2M samples, which advances M samples between adjacent

transforms. It is critically sampled and only M coefficients

are generated for each 2M set of input samples. Compu-

tationally efficient implementations have contributed to

the widespread use of the MDCT in many audio coding

standards.

Figure 12 shows a generic block diagram of an au-

dio encoder incorporating the filter bank and the percep-

tual model. The resulting spectral components (MDCT

coefficients) are quantized in such a way that the result-

ing quantization noise is inaudible. This is accomplished

by using the masking level obtained from theperceptual

model.The amount of noise is controlled by the resolu-

tion of the quantizer. By choosing a different quantizer

step size the amount of noise can be quantized. Typi-

cally the same quantizer is used for a set of coefficients,

and the corresponding step size for that set is transmit-

ted to the decoder. It should be noted that at this point,

due to quantization, the decoded signal will be different

from the original (i.e., lossy coding). To accomplish per-

ceptually lossless coding, we need to make sure that the

quantization noise remains below the masked threshold.

The redundancy removal is accomplished by encoding the

quantizer indices with lossless coding techniques (e.g.,

Huffman coding). To avoid confusion it should be clear

that this is a lossless coding technique on quantizer in-

dices; hence the overall operation is still a lossy coding

Analysis

Filter

Bank

Analysis

Filter

Bank

QuantizerQuantizer Noiseless

Coding

Noiseless

Coding BitstreamFormatter

Bitstream

Formatter

Perceptual

Model

Perceptual

Model

Noise

Allocation

Noise

Allocation

audio

input

IRRELEVANCY REMOVAL REDUNDANCY

REMOVAL

Figure 12: Block diagram of generic audio encoder.