Index |
Fishing in the noise (published in Electron #7, 2007) Introduction For many years radio-amateur are trying to
communicate especially under adverse contact conditions. A dedicated
radio-amateur is most satisfied when he finally does succeed in 'working that
difficult station' when conditions are almost prohibitive and 'big guns' have
given up long ago. The limitation of the license than no longer will be
regarded as a threat but more of an opportunity and a challenge to use every
trick in the book to bring the minimal exchange of information to a good end. The challenge is taken on at various levels,
starting off at the antenna at the transmitting and receiving side. Loss in
the antenna system is the first to avoid trap and a discussion about this
subject may be found at "Where
does the power go". Next to the antenna-system, the 'art of
listening' is an important instrument that may only be mastered after
extensive practice. This art may be boost by clever tools to give the signal
an edge over the noise. This last subject will be the topic in the following
discussion. Already from the beginning of
radio-communication, listening stations have been confronted with all kinds
of natural radio-sources that made life difficult. It therefore did not take
long before the 'noise limiter' was invented to limit the deafening burst
noise of thunderstorm. In these early radio-years Morse code signals
were the mode of operation. Using feed-back receivers it was possible to
limit band-width at the receiving side to the absolute minimum as required
range for these signals.
Frequency-stability problems of the receivers of those days however made
keeping track a tedious adventure. With the super-heterodyne arriving,
stability was enhanced as well and filtering to improve signal to noise ratio
was shifted to the audio part of the station. It soon appeared to be difficult to design
and construct good audio-filtering that was steep enough to prevent unwanted
signals to pass without too much added distortion. Filtering technology at
the IF section was helping out especially with the aid of 'mechanical'
technologies like crystal filters, later to be followed by the famous Collins
mechanical filters. With these technologies firmly established it
appeared rather difficult to further improve on analogue filtering since
practical limitation proved to be of a rather fundamental nature. Listening
practice was again the art to go for. Digital signal processing A big step in filtering technology could be
made with the micro-processor arriving on the scene to allow Digital Signal
Processing (DSP) on real-time signals. By converting analogue signals to digital
values using an A/D converter, mathematical signal processing became
available with addition, subtraction, multiplication and holding the signal
samples for some time (delay) as tools of the trade. Loss of signal quality was related to limited
resolution of readily available, real-time A/D converters and the span of the
internal registers. In those early processor days these internal registers
were four to eight bit wide while the processor was running at a very modest
speed of a few mega 'beats' per second. To reliably reproduce digital
signals, the signal samples should be taken at just over twice the highest
frequency to reproduce. With the high number of clock periods for each
mathematical process taken into account it is clear the micro-processors of
those day where not fast enough for real-time audio processing. By the end of the '80 of last century the
first micro-processors came to the market that were dedicated to real-time,
digital signal processing. Registers were 16 to 24 bits wide with on-board
high-resolution A/D and D/A converters to provide for the necessary high
signal resolution. Furthermore these processors were equipped with dedicated
hardware for direct multiplication without the high amount of clock tics. The TMS 310 by Texas Instruments was a good example
of such micro-processor and is still at the heart of many DSP processing
equipment, although currently in more advanced versions of this TMS
series. In many contemporary receiving systems DSP is
'a must' and a selling argument. Next to build-in processors add-on DSP
systems are on offer by various
manufacturers as a 'set-top' box. More information on this subject is also
available at the ARRL-site.
Many of these DSP processors are limited to
analogue-like filtering, although with much enhanced specifications and
without the 'ringing'- effects that hinder their analogue counterparts when
at low band-width. With these DSP filters it is not unusual to find very
steep filter edges e.g. - 3dB @ 2,5 kHz with -60 dB @ 2,8 kHz. It is also
possible to select variable (audio) bandwidth with these specifications at
any position in the LF spectrum. With this technology also (automatic) notch
filters can be created to counteract annoying whistlers. Most of the above functionality however is
still connected to processing in the 'frequency domain', whereas with DSP
technology and the right mathematics also processing in time-domain is quite
feasible and also time and frequency domain at the same time. Mathematical
background to this extend has become available from the military and medical
profession opening doors to more, and more versatile audio signal processing. A good example of such a dedicated DSP system
is the 'SprachEkstraktor' by Günther
Michels, DJ7UP, available as module, DX-11 or as a
complete set-top box, DX-21 and may be found at his web-site. This system is exploiting the typical
characteristics of spoken language as in the limited frequency range and the
low frequency repetition of basis speech utterances. By focusing on these
characteristics the system is capable of recognizing speech from (other?)
noises and enhancing accordingly. This typically is also the power of our
biological hearing and speech processing.
Not long ago I visited a local ham-friend
that showed me the DX-21 'set-top box'. It indeed did sound OK and was
remarkable in its ability to remove the noise from already intelligible
speech signals. In my experience 'speech enhancing' systems can improve the
quality of speech signals in a noisy environment and will diminish listening
'fatigueness'. The biological processor though is
usually better at improving intelligibility at adverse signal conditions, so
to me it was not really worth my while to purchase such system. This speech processing system by DJ7UP
however appeared to have opened a new window and I was wondering whether this
system really was to improve on an already trained biological processor.
Therefore I acquired the bare processing board DX-11 to test for myself. The DX-11 kit as send to me was existing of a
professional multilayer printed circuit board, equipped with SMD components
amongst which the earlier mentioned TMS320 in a prominent position. On top of
this a small amount of additional components was supplied as well to ease
integrating this processing system in an existing environment (receiver,
speaker) or to construct your own stand-alone cabinet. A picture of the kit
is in figure 1. Figure 1: DX-11 kit The dimensions of the small footprint DX-11
board may be judged by the size of the additional components that came with
the board. It is obvious this small PCD will easily fit in any, not too
compactly build, existing cabinet. To enable an objective, quantifiable and
repeatable test, I set up a test environment as in figure 2.
In figure 2 the following components may be
recognized: 1. At position 2. Amateur receiver set for reception of SSB
speech signals. Internal audio bandwidth is switched to 2,4 kHz @ -3dB and
4,4 dB @ -60 dB. This receiver was to be tuned to
different types of band-noise, switching noise, TV-rattle, thunderstorm,
etc. 3. Audio filter to limit the 'perfect speech'
source to the same audio band bandwidth as the receiver. The signals from 1
and 2 are added and fed to the DX-11 at position 5. 4. The input signals to the DX-11 are also
fed to an oscilloscope to ensure peak signals are below the maximum input
voltage for the A/D converter in the processor at 900 mVtt.
At the same position an additional, slow reacting, analogue rms milliVolt meter is
connected to obtain a clear and repeatable reading of all input signals as a
single value. 5. DX-11 that could be controlled by the
supplied potentiometer to obtain optimal intelligibility at different
noise-types and signal-to-noise conditions. A switch has been inserted to select signals
with and without the speech-processor. The final audio signals were
level-optimized each time for good listening conditions. This appeared to be
a little bit different for different types of signal. For each test the maximum levels have been
checked and set to ensure maximum peak-peak input signal always remain (just)
below 900 mV. The accompanying rms value has been
kept constant at each of the successive tests, to ensure the DX-11 to always
operate at the same internal resolution. The first series of speech intelligibility
tests have been performed using back-ground band noise from all HF amateur
frequency bands, because this noise did sound differently in different bands.
Next, speech was turned up until this was just, but 100% intelligible. At
each of these test the optimal setting at the DX-11 potentiometer was
'tweaked', but this appeared not to be critical; once a setting had been
selected, this could be maintained for most of the other tests. At this 'just 100% intelligibility' position,
the noise source (receiver) was switched off, and the speech level was
measured. This speech level as compared to the noise measurement above
(without speech) is providing the S/N figure at this setting. The test could
be repeated (with different speech content, the news or spoken book was not
interrupted) with very good accuracy.
Next, the test was repeated without the
speech processor, to find the set-point for the trained, naked ear, at the
same audio noise level as before. Again the 'just 100% intelligibility' position is being
determined to yield a comparable S/N figure. Comparing both S/N figures will present an
'objective' presentation as to the quality (improvement / degradation) of
this speech processor. Tests have been performed for various types
of 'noise', e.g. 'white' band noise, noise by TV-rattle, jammers,
thunderstorm, LF break-in noise, neighboring channel SSB-speech break-in and
artificially made 'noise'. For all measurements, different 'voices' have been
used, as some were more easily intelligibility or pleasant to the ear. In
general, the more high-pitched voices (female), at a more constant speech
rate (female) provided highest intelligible in noisy conditions when
operating the DX-11 as well as to the naked ear. Results - White band-noise at a 'clean' part of the * With processor: S/N: -9,3 dB. Without: S/N: -6,3 dB.
DX-11 improves the intelligibility by 3 dB. Some difference could be measured between
various male voices with female voice as best. Results at this voice: * With processor: S/N: -10,9 dB. Without: S/N: -6,3. DX-11 appears to be better tuned
to this voice. On average DX-11 improved S/N by around 4 dB. - White band noise in the * With processor: S/N: -15,3 dB. Without: S/N: -9,3 dB. It is remarkable we can hear 'deeper' into
the noise with this type of 'colored noise' than without. Also DX-11 is
better at this task. In all, DX-11 was beating 'naked hearing' by 6 dB, an
improvement of 2 dB over 'band noise' only. Since not all 'TV-rattle' sounds the same, a
different and much stronger rattle at + 3 S-points was found at the * With processor: S/N: -8,6 dB. Without: S/N: -2,3 dB. This rattle appeared to be a challenge for
DX-11 as well as to the naked ear. Still DX-11 won by about 6 dB. - Same band, same band-noise, this time with
added artificial rattle by a LF square wave generator, tuned to a spike
repetition of about 5 Hz. (at the human utterance-repetition frequency
range). * With processor: S/N: -6,8 dB. Without: -4.0 dB. The DX-11 as well as the naked ear clearly
had some problems. Still DX-11 won by 3 dB. - Wide band noise at a 'clean' part in the * With processor: S/N: -6,3 dB. Without: S/N: 4,0 dB. Noise was more 'white' to the ear with more
signal-to-noise needed for full intelligibility. Still DX beat the naked ear
by over 2 dB. - Wide band noise on 7,05 MHz.
with voice break-in by next-door station (side-band whisper). Processor nor naked ear could improve on this
situation, probably because this 'noise' is also of a voice-type character.
This will also happen when trying to break through a (voice) pile-up. Above tests and measurements are leading to a
few interesting observations: - A (trained!) human ear is capable of
recognizing speech signals that are up to 6 dB below the noise; this may even
go up to 9 dB depending on the particular type of noise. - DX-11 will improve on the naked, but
trained ear by 2,5 - 4 dB depending on noise 'coloring' and voice type. - At a particular type of 'jamming', DX-11
may even improve up to 6 dB. - If 'jamming' is at the speech repetition
rate ( 5 - 10 Hz.), DX-11 improvement is lowering to around 2,5 dB. - At 'noise signals' that are speech-related
(side channel break-in), no improvement may be noticed. - At already well intelligible speech
signals, DX-11 will further improve speech 'contrast' by lowering the noise
considerably. This will make listening to these signals less tiresome. This
type of speech enhancement therefore may also be useful in other noisy
environments, e.g. communicating at a motorbike - Interesting to notice: DX-11 will 'improve'
intelligibility of the lead-vocal in an orchestra; apparently the orchestra
is also regarded as 'noise' by the DX-11. - The DX-11 does not improve for other than
speech type of signals (e.g. Morse-code, other modes). Apparently the system
was not designed for this. General conclusion is justified the DX-11 can
make a difference when trying to (voice) communicate under adverse signal to
noise conditions. The improvement is in the order of 3 - 6 dB S/N over the
naked, but trained ear. With speech signals already intelligible, DX-11 will
further improve quality by diminishing background noise. To obtain the same
improvement without the DX-11, the antenna should be at least twice as
effective in improving the signal or the transmitter should crack-up output
power by a factor of 2 - The test results were convincing enough to me
to finally build this DX-11 into my ham-bus system,
you may also find at this web-site. Bob J. van Donselaar, on9cvd@veron.nl |
|