The Internet Encyclopedia (Volume 3)

P1: C-172

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

308 SPEECH ANDAUDIOCOMPRESSION

Frequency Range (Hz)

10 20 50 200 3400 7000 15k 20k

TELEPHONE

WIDEBAND

FM-QUALITY (mono)

COMPACT DISC (mono)

kb/s

705.6

512

256

64

Figure 1: Relationship between bandwidth and bit rate.

decoderor stored for later retrieval. The decoder takes the

bit stream and generates the corresponding decoded PCM

signal, which is a rendering of the original input PCM

signal.

In the remainder of this chapter we will focus on reduc-

ing the bit rates of speech and audio signals while provid-

ing the best possible signal quality. At this point it is good

to point out that quality is a difficult attribute to quantify

because it has many dimensions. Most of these are asso-

ciated with specific applications, and it is important to set

proper objectives when designing or choosing a particular

compression algorithm. For example, for speech commu-

nications it could be important to have consistent perfor-

mance for various input conditions, such as clean speech,

noisy speech, and input level variations. For compression

of music signals it could be important to have consistent

performance for various types of music, or to preserve au-

dio bandwidth and stereo image as much as possible. As

will be clear later, this quality objective will be constrained

by other factors such asdelay(the time needed to encode

and decode a signal) and thecomplexity(the number of

arithmetic operations) of the methods used.

Speech and audio compression applications can be

divided into two classes. The first isbroadcasting(e.g.,

streaming); the other iscommunication(e.g., Internet

telephony). Each application has different requirements,

as can be seen from Table 1.

There are two principal approaches to the compression

of digital signals:losslessandlossycompression. Loss-

less compression techniques take advantage of redundan-

cies in the numerical representation. For example, instead

of using 16 bits per sample uniformly, one could use a

new mapping that assigns symbols of shorter length to

the most frequent values. Or if the signal values change

slowly between sample values, one could encode the dif-

ferences instead, using fewer bits. Lossless compression

is a reversible operation and the input and output signal

samples of Figure 2 will be identical. Lossy compression

techniques, on the other hand, assume that the signal has

encoder transmission

or storage

PCMinput decoder PCMoutput

bitstream bitstream

Figure 2: Block diagram of a generic coding or compression

operation.

Table 1Difference in Requirements for Broadcasting

and Communications

Application Broadcasting Communications

Characteristic One-way Two-way

transmission transmission

Delay Not important Important

Complexity Not important Important

Technology Audio coding Speech coding

a human destination, which means that signal distortions

can be introduced as long as the listener either is not able

to hear them, or has no serious objections. Lossy com-

pression is not reversible and the input and output signal

samples of Figure 2 will be different. For many signals

this difference would be unacceptable, but for audio sig-

nals, one only worries about audible differences. If the

differences are inaudible the lossy coding techniques used

are often referred to asperceptually losslesscoding tech-

niques. But even if the differences were clearly audible it

still would be acceptable for many applications. A good

example is the difference between telephone speech and

natural speech, where the telephone signal is significantly

limited in bandwidth (typically less than 4 kHz).

In practice, the use of lossless coding for audio and

speech signals results in limited compression efficiency

and its use is restricted to high-quality applications such

as the audio DVD. The compression efficiency for lossy

coding can be significantly higher and consequently per-

ceptually lossless and lossy coding are the main ap-

proaches used in most audio and speech compression ap-

plications. Speech compression takes this approach one

step further by also assuming that humans generate the

source signal, which gives it certain properties that can

be taken advantage of by the compression algorithms. As

a result, speech compression can achieve very high com-

pression efficiency. Figure 3 gives a summary of the typical

bit rates and applications.

Compression for Packet Networks

The compression algorithms described in this chapter are

used for communication and broadcasting over wired and

1 2 4 8 16 32 64 128 1 2 4

kb/s Mb/s

secure voice

wireless

voice messaging

network telephony

video conferencing

audio broadcasting

Quality

audio storage

CD quality

FM quality

wideband quality

telephone quality

Figure 3: Typical bit rates and applications for speech and

audio compression.