494 Chapter 15
quantization noise over four times the spectrum, thus only 1/4 of the noise power now
resides in the audio band. If we assume that the noise spectrum is uniform and that
dither has been judiciously applied, this is equivalent to obtaining a 1-bit enhancement
in dynamic range within the audio bandwidth in either digital or analogue domains. This
performance can be further improved by the process of noise shaping.
The information capacity of a communication channel is a function of the SNR and
the available bandwidth. Thus there is room for trading one against the other. The
higher the oversampling ratio, the wider the bandwidth in which there is no useful
information. If samples were to be coarsely quantized it should be possible to place the
extra quantization noise in part of the redundant spectrum. The process of relocating the
noise in the redundant spectrum and out of the audio band is known as noise shaping
and it is accomplished using a recursive fi lter structure (the signal is fed back to the
fi lter). Practical noise shaping fi lters are high order structures that incorporate integrator
elements in a feedback path along with necessary stabilization.
The process of oversampling and noise shaping can be taken to an extreme, and
implementation of this approach is available in certain DACs for compact disc systems.
The audio has been oversampled by 256 in the Philips bit stream device, 758 in the
Technics’ MASH device, and 1024 in Sony’s device. The output is in the form of a
pulse train modulated by its density (PDM), by its width (PWM), or by its length (PLM).
High oversampling ratios are also used in ADCs, which are starting to appear on the
market at the current time.
15.11 Transform and Masking Coders ............................................................................
We indicated very early on that there may be some advantage in terms of the reduction
in data rate to taking the Fourier transform of a block of audio data and transmitting the
coeffi cient data. The use of a technique known as the discrete cosine transform is similar
in concept and is used in the AC-2 system designed by Dolby Laboratories. This system
can produce a high-quality audio signal with 128 kb per channel.
The MUSICAM process also relies on a model of the human ear’s masking processes.
The encoder receives a stream of conventionally encoded PCM samples, which are then
divided into 32 narrow bands by fi ltering. The allocation of the auditive signifi cance of
the contribution that each band can make to the overall program is then carried out prior