The Internet Encyclopedia (Volume 3)

P1: C-172

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

SPEECHCODINGTECHNIQUES 313

ENCODER

DECODER

A(z) P(z) quantizer

Inversequantizer 1/P(z) 1/A(z)

input bits

bits output

Figure 9: Block diagram of a linear predictive coder

with short-term and long-term prediction.

This is achieved by having a number of possible rep-

resentative sample sets orcodebooks. Quantization is

achieved by selecting the codebook entry (each entry con-

taining multiple samples) that is the best representation

of the signal to be quantized. The codebook index cor-

responding to this entry is transmitted and the decoder,

which has a similar codebook, uses this index to look up

the corresponding values. Whereas the lowest quantiza-

tion rate for scalar quantization is 1 bit per sample, VQ

allows fractional bit rates. For example, quantizing 2 sam-

ples simultaneously using a 1-bit codebook will result in

0.5 bits per sample. More typical values are a 10-bit code-

book with codebook vectors of dimension 40, resulting in

0.25 bits/sample. Both scalar and vector quantization at-

tempt to find the quantized value that is the closest to the

original unquantized input. In the block diagram of

Figure 9 one can argue that the overall goal is not to find

the best quantized residual values, but the best quantized

speech signal. Especially for coarse quantization (very few

bits per sample), this becomes an issue. A powerful tech-

nique used in speech coding isanalysis-by-synthesis,in

which the effect of quantization is determined by examin-

ing the effect on the decoded output. This is accomplished

by operating the diagram of Figure 9 in the configuration

shown in Figure 10.

In this figure the encoder of Figure 9 is enhanced with

a local decoder. In most practical coders the predictors

are still computed as usual. The quantization of the resid-

ual signal is done in an analysis-by-synthesis fashion. If

we assume that we use a codebook, then the essence of

this approach requires that for each codebook vector we

perform local decoding and compare the resulting pro-

totype output with the original input signal. The code-

book vector that gives the best approximation is selected.

Note that with this paradigm we are indirectly creating a

quantized residual signal, and this signal is often referred

coder decoder +

error

minimization

error

weighting

• coder decoder

error

minimization

error

weighting

Figure 10: Analysis-by-synthesis encoder with

error weighting.

to as anexcitationsignal. Instead of directly comparing

the original input signal and the quantized and decoded

rendering of this signal, an error-weighting filter is intro-

duced, which better reflects the way our auditory system

perceives distortions. It should be noted that this weight-

ing is much simpler than the complex auditory models

used in audio coding. The block diagram of Figure 10

forms the basis for a family of coders generically referred

to as code-excited linear predictive (CELP) coders. Most

modern speech-coding standards are based on this princi-

ple. A large amount of research has been done on efficient

codebook structures, which not only give the best perfor-

mance, but are also manageable in terms of search. A com-

monly used structure is the so-calledalgebraiccodebook,

which consists of a few nonzero pulses with deterministic

positions.

For all coders several techniques can be used to fur-

ther improve their performance. Some of these tech-

niques can be done independent of the coder, although

in practice it makes sense to take advantage of the pa-

rameters already computed by the coder. Preprocessing

techniques that are useful are gain control and noise

suppression. The latter can be quite sophisticated but

very effective, especially for the lower rate speech coders,

which typically do not handle background noise well.

A widely used form of postprocessing ispostfiltering.In

this process the decoded speech signal is slightly dis-

torted in such a way that the coding noise gets suppressed

and the signal gets enhanced. If done with care it can

clean up a signal, resulting in perceived quality improve-

ment.

Another technique that has found some popularity is

taking advantage of the fact that conversational speech

comes in bursts, due to the speakers talking at alternate

times. Sometimes there can be large pauses in between

words. This can be taken advantage of by only transmit-

ting when active speech is present. When speech is not ac-

tive no signal is transmitted. Because on the average peo-

ple speak half of the time, this technique has the potential

to reduce the bit rate by half. To make thisdiscontinu-

ous transmissionapproach work, avoice activity detector

(VAD) is needed. For speech without background noise,

this approach can work quite well. When background

noise is present (e.g., a car), it is more difficult to get reli-

able decisions from the VAD. Moreover, when no talker is

active, and no signal is transmitted, the receiver side needs

to substitute a replacement signal. This is referred to as

comfort noise.For high levels of background noise it is

difficult to have this comfort noise match its characteris-

tics. Hence more sophisticated systems transmit low-rate

information about the background noise, such as energy

and spectral characteristics, at average rates of about 1 to

2 kb/s.

Speech Coding Standards

For communication purposes it is important to es-

tablish standards to guarantee interoperability between

equipment from different vendors, or between telecom-

munication services in different geographic areas.

Telecommunication standards are set by different stan-

dard bodies, which typically govern different fields of use.