[RTTY] 24 bit sound card

[Kok Chen]

On Mar 18, 2005, at 6:32 AM, David Hachadorian wrote:
> Is there any advantage to using a 24 bit sound card when
> demodulating rtty with MMTTY?

I'll bet that you will find that there are various opinions on this.  
Here are my 2 cents.

Warning, this is long.  If you don't want to read through all of it, 
the gist is that IMHO, it depends on whether you use narrow IF filters 
or wide IF filters.  If there is there is no interfering signal in the 
passband, then a 16-bit converter is ample if you are willing to ride 
the audio/RF gain.  If you have the possibility of two signals in the 
same passband or you are using "wide passband RTTY", then the analysis 
becomes more complex.  It also depends on whether your operating 
system's audio library can support anything more than 16 bits -- the 
MacOS X's audio framework is floating point, for example.

First consider a 16 bit converter.  The maximum numerical range is of 
the order of 96 dB assuming that 1) the A/D converter is linear to 1 
out of 1 bits, 2) there is some noise dithering and 3) the noise floor 
of the stages that precede the is low enough to support the 96 dB.

However, this numerical range is not quite the same in some sense as 
the dynamic range.  Let us say there are two signals in the audio 
passband.  If the loud signal is 96 dB louder than the weaker signal, 
the weaker signal is itself a 1 bit (square wave) signal.  There is 
something in the Radar Astronomy world called the Van Vleck theorem (I 
think it is also called the "arc-sine rule" in Probabililty theory, but 
it has been a long time sine I cracked open a volume of feller or 
Papoulis, so don;t hold me to this).  What the theorem says is that you 
can recover the spectrum of a hard clipped signal perfectly, if it is 
in a strong Gaussian noise environment, but at a loss of 2.19 dB of 
SNR.

At the threshold of FSK detection, 2 dB of SNR change (for the case 
where there is no QSB, just noise) is huge, and can be a difference of 
practically printing everything (one or two bad prints per line) to 
printing lots of garbage (one out of 4 characters are bad).  Take a 
look at the plots in textbooks and the AWGN plots that Alex VE3NEA had 
posted a month or two back on this reflector.  So, 2 dB lowered SNR is 
a "bad thing" for us RTTY folks.

I had done some work many eons ago to extend van Vleck to the multi-bit 
case (playing with joint Gaussian distributions and integrations 
thereof) and showed that the SNR ratio loss can be made "negligible" 
when you are up at the 4 bit region instead of the 1 bit (hard clipped) 
region.  In practice, we don't have Gaussian noise on the ham bands, 
and no one uses the arc sine transform to receiver spectra, so we will 
do worse that "2.19 dB" of SNR loss.

So, the 96 dB numerical range of a perfect 16 bit converter now reduces 
to 72 dB of usable dynamic range if you need 4 bits to represent the 
weaker signal.

Is 72 dB enough?

First, the 72 dB is for ideal additive Gaussian noise and if you also 
apply the corrective transformations.  My guess is that in practice, 
you are playing more in the area of 60 dB to 70 dB.

Now lets consider two cases: 1) you are using a 250 Hz filter and 2) 
you are using  a 2500 Hz filter ahead of the A/D converter.

With a 250 Hz filter, unless another signal straddles the desired 
signal, you are working with a single signal case.  I cannot imagine, 
in this case that you need more than 30 or 40 dB of dynamic range, as 
long as there is no QSB which your rig's AGC has not taken care of, and 
the band conditions don't change over time.  Afterall, under the single 
signal case, even a hard clipped (some people call it limitered or "FM" 
demodulation of RTTY) signal prints pretty well, although both RITTY 
and cocoaModem use linear matched filters and work better when there is 
no limiter or hard clipping whatsoever.

Furthermore, band conditions could change so you might have to ride the 
gain of your transceiver so that the A/D converter "sees" a signal that 
neither clips it nor less than 4 to 6 bits in magnitude -- this is 
where more bits help.   Modems such as the ST-8000 (85 dB dynamic 
range) and the Timewave DSP-599zx have built in AGC.    Notice that 
good 24 bit converters have between 95 to 105 dB weighted noise floors, 
and they can basically give you better dynamic range than modems of the 
past which had to use AGC!  AGC has its own problems -- you may have 
noticed that with AGC engaged in a modem, you might miss the first 
couple of characters when a very loud CQ'ing station starts up in a 
contest.

If you are willing to install a pot before the A/D converter (or have a 
mechanism to adjust the attenuation in the A/D converter) and have a 
way to constantly monitor the signal to make sure the signal is large 
enough but not saturating the A/D converter, then a 16 bit converter, 
with the operator manually riding the gain is plenty good enough, IMHO 
in the single signal case.

However, the problem is different if you do any "wide passband RTTY," 
i.e., watching an entire 2 to 3 kHz, or more passband).  In this case, 
you just about always have more than two signals in the passband.  
Under a DX pileup condition (where you can see both DX and the pile) or 
under contest conditions, the "60 dB to 70 dB" worth of dynamic range 
is not good enough for a wide passband.  That is assuming that the 
rig's front end dynamic range is good at least to the 70-80 dB region 
in the first place.

My crazy prediction is this:  RTTY will eventually move to use wide 
passband and perhaps also the point and click waterfall tuning as 
practiced today in PSK31.  There are simply too many advantages with 
wide passband operation (for a DX station picking a signal from the 
pile, for a non-DX station fining where in the pile to call, for 
multiple decoding streams in a contest when you are doing S&P, etc etc.

To use wide passband RTTY, you will need A/D converters that have 
better dynamic range than the rig's front end or the A/D converter will 
be limiting the overall dynamic range of your station.

Personally, I use a M-Audio Transit (24-bit) as my A/D converter, into 
a homebrew software modem that does make use of 24 bits of numerical 
range.  Your mileage may vary.

And be careful... not all "24-bit" converters perform the same, just as 
not all "16-bit" converters perform the same.  Look for actual dynamic 
range, including noise floor contributions, and not just count the 
bits.

73
Chen, W7AY