This is only a preview of the May 1990 issue of Silicon Chip. You can view 44 of the 104 pages in the full issue, including the advertisments. For full access, purchase the issue for $10.00 or subscribe for access to the latest issues. Articles in this series:
Articles in this series:
Articles in this series:
Articles in this series:
Articles in this series:
|
A new error correction
technique for digital recording
All digital recording and playback techniques have
some form of error correction which allows signals
to he restored after transmission errors. But really
large errors present a problem in that they can
practically destroy the signal. Now, a new
technique has been devised with offers significant
improvements.
At the Philips Research
Laboratories in Eindhoven, Raymond Veldhuis has developed a
method which allows even large errors in signals, such as speech,
music or images, to be restored. For
this he makes use of the regularity
present in almost every signal. By
analysing the error's environment,
it is possible to correct the error in
such a way that there is no longer
any perceptible distortion or picture disturbance.
Us1ng knowledge of human hearing, Veldhuis also succeeded in
reducing the bandwidth required
for the transmission of audio
signals by a factor of between six
and eight without any audible loss
of sound quality.
This means that the bandwidth of
an FM transmitter can be quite sufficient for the transmission of
digital music. The coding method
for this bandwidth reduction has
been included as a proposal in the
Eureka project on Digital Audio
Broadcasting (DAB).
Digitisation
Audio and video signals as we
perceive them are analog. Audio
signals vary continuously over time;
picture signals change continuously
as a function of place and time. The
continuously changing value of, for
example, sound pressure, pitch or
picture brightness can be recorded
(digitally) with a high level of accuracy by taking a sample of the
signal at very specific time
intervals.
For example, in the case of
signals recorded on compact disc,
the signal is sampled 44,100 times
per second. The sampling frequency is then specified as 44. lkHz.
The digitised signal is a series of
binary numbers; ie, zeros and ones
which are also called bits. If there
is a fault, then a bit error can arise
with the result that a binary
number is no longer completely correct. When the fault is not too
large, such errors can be both
detected and corrected by adding
extra bits to the signal. However, in
the case of large errors, such as the
loss of several milliseconds of a
conversation on a earphone due to
interfering reflections from mountains or high buildings, other ways
of restoring the digitised signal
must be sought.
The environment
Veldhuis bases his restoration
method on the fact that speech,
music or picture signals all have a
certain regularity, characterised
by the signal spectrum.
This regularity can be measured
and then it is possible to replace the
missing numbers in the series so
that the restored part of the signal
shows, as far as possible, the same
regularity as the rest of the signal.
Speech signals
........ ....
,
~~~········
••••••
~~L
·••
If the data stream for a picture is corrupted, lots of picture elements will be
lost, as shown in this extreme case.
4
SILICON CHIP
When applied to speech signals,
the technique has been used to
restore gaps as long as 12.5
milliseconds. When the restored
signals were observed on an
oscillograph, the differences between them and the original signals
were negligible. And in listening
tests, no difference could be heard
between the original and the
restored signal.
The speech sampling frequency
in the case of digital telephony, as
used in car telephones, is 8kHz
...
This photo shows how the new error correction method can restore the
seriously corrupted picture on the opposite page. It can also restore badly
corr.ected data streams for audio signals.
(8000 samples per second). Hence,
around 100 consecutive samples
were restored in the 12.5 millisecond section of signal.
Music signals
In the restoration of music
signals, Veldhuis also used, in addition to the regularity concept, the
fact that a music signal can be considered a s the output signal of a
filter. The spectrum is then fully
characterised by the filter coefficients. It is thus possible to devise a
mathematical equation with which
every unknown sample can be
estimated on the basis of the
preceding samples.
If some of the samples following
the unknown sample are also
known, then a prediction can be
made which comes very close to the
original sample. In practice, it is
possible, at a sampling frequency of
44. lkHz, to fill in up to 30 successive missing samples by calculation, without any musical defeqts
being apparent.
Picture signal correction
Picture signals can also be
digitised. Still pictures can be conceived as a changing brightness
signal varying according to the
position (ie, in two dimensions).
With digitisation, the signal be-
comes a field of numbers. In such
a field, every number represents
the brightness of a picture element
[pixel). Moving pictures can then be
considered as a succession of still
pictures.
In the (future) transmission of
digitised pictures, for example,
groups of 8 x 8 pixels will be
transmitted. If there is a transmission error then an entire 8 x 8 area
can suddenly disappear and this is
seen as a picture fault. Restoration
can again take place in the manner
indicated; determine the regularity
in the brightness distribution in the
area's surroundings and from this,
calculate the brightness of the missing picture elements.
As the accompanying photos
demonstrate, a picture full of
transmission errors can be fully
restored - a dramatic demonstration of the technique.
Economical coding
The digitisation of signals,
whether they be music, speech or
picture signals, gives an excellent
quality of reproduction and, as indicated above, possibilities for effectively repairing damaged signals. There is, however, a price to
pay; the sampling frequency required for good quality must be at
least twice as high as the highest
frequency to be reproduced. In addition, a number of bits are needed
to record the content of a sample
digitally.
An audio signal with a maximum
frequency of Z0kHz calls for a
sampling frequency of 44. lkHz.
Further, to digitally code each sample, an accuracy of 16 bits is required in order to prevent the
disturbing influence of rounding errors. In this way, the reproduction
of a signal with a bandwidth of
Z0kHz requires a bit frequency of
700,000 bits/second; for a stereo
signal this figure has to be doubled
again.
The resulting requirement of 1.4
million bits per second far exceeds
the capacity of an FM channel,
making economies necessary in the
coding of the signals. However,
with data compression, the bandwidth required for the transport of
digital signals can be greatly
reduced.
Signal masking
Studies of hearing have shown
that strong signals with a certain
frequency mask weaker signals
with neighbouring frequencies; ie,
make them inaudible. This only happens if these neighbouring frequencies do not differ too greatly from
the frequency of the stronger
signal, and their strength does not
exceed a certain threshold value
(the masking threshold).
However, if a musical signal is
divided into narrow frequency
bands, then it is possible to make do
with a rougher coding (less bits per
sample) as the resulting interference then remains below the
masking threshold.
In cooperation with the IRT, the
German Institute for Broadcast
Technology, and the CCETT, the
laboratory of the French Post and
Telecommunications organisation,
this coding method has now been incorporated as a proposal in EUREKA
project 147, Digital Audio Broadcasting. The purpose of this EUREKA
project is to arrive at a new
transmission standard for digital
audio.
Footnote: the results described
here relate exclusively to laboratory research. They do not involve
the marketing or manufacturing of
new products.
~
MAY1990
5
|