JT9 is a mode designed for amateur QSOs at MF and LF.  The mode uses
the same 72-bit structured messages as JT65.  Error control coding
(ECC) uses a strong convolutional code with constraint length K=32,
rate r=1/2, and a zero tail, leading to an encoded message length of
(72+31)*2 = 206 information-carrying bits.  Modulation is 9-FSK: 8
tones for data, one for synchronization.  Sixteen symbol intervals are
used for synchronization, so a transmission requires a total of 207/3
+ 16 = 85 channel symbols.  Symbol durations tsym are approximately
(TRperiod-8)/85, where TRperiod is the T/R sequence length in seconds.
Exact symbol lengths are chosen so that nsps, the number of samples
per symbol (at 12000 samples per second) is a number with no prime
factor greater than 7.  This choice makes for efficient FFTs.  Tone
spacing of the 9-FSK modulation is df=1/tsym=12000/nsps, equal to the
keying rate.  The total occupied bandwidth is 9*df.  The generated
signal has continuous phase, and there are no key clicks.

Parameters of five JT9 sub-modes are summarized in the following
table, along with S/N thresholds measured by simulation on an AWGN
channel.  Numbers following "JT9-" in the sub-mode names specify the
T/R sequence length in minutes.

--------------------------------------------------------------------------
Mode     nsps nsps2  df    tsym  BW   S/N*   Tdec Tfree Factors 
        12000  1500 (Hz)   (s)  (Hz)  (dB)   (s)   (s)  of nsps      nfft3
--------------------------------------------------------------------------
JT9-1    6912   864 1.736  0.58 15.6 -26.9   52.5  7.5 2^8 3^3        2048
JT9-2   15360  1920 0.781  1.28  7.0 -30.2  112.3  7.7 2^10 3 5       2048
JT9-5   40960  5120 0.293  3.41  2.6 -34.4  293.6  6.4 2^13 5         6144
JT9-10  82944 10368 0.145  6.91  1.3 -37.5  591.0  9.0 2^10 3^4      12288
JT9-30 252000 31500 0.048 21.00  0.4 -42.3 1788.5 11.5 2^5 3^2 5^3 7 32768
--------------------------------------------------------------------------
* Noise power measured in a 2500 Hz bandwidth.
NB: nfft3 might be doubled and used with a sin^2 window.

Transmitting
------------
1. Source encode the structured message to 72 bits
2. Apply convolutional ECC (K=32, r=1/2) to yield (72+31)*2 = 206 bits
3. Interleave to scramble the bit order
4. Assemble 3-bit groups to make (206+1)/3 = 69 symbols
5. Gray-code the symbol values
6. Insert 16 sync symbols ==> 69+16=85 channel symbols, values 0-8


Receiving
---------
1.  Apply noise blanking with the timf2 method
2.  Filter to 1000 Hz bandwidth and downsample (1/8) to 1500 Hz, saving
    complex data to array c0(2,700,000).
3.  Compute spectra at half-symbol steps.  Use for waterfall display 
    s(22000) and save in ss(184,22000) and savg(22000) for detecting 
    sync vectors.
4.  At time Tdec, find sync vectors in ss(); get approx DF or list of DFs
5.  Do full-length FFT, NFFT1=96*nsps2, zero-padded as required.
6.  For each candidate signal, do inverse FFT of length 1536 (or 3072?).  
    This yields 16 complex samples per symbol; sync tone should be 
    close to zero frequency.
7.  Use afc65b method to get improved values of DF, DT.
8.  Tweak freq and time offset to 0.
9.  Compute 8-bin spectra of 69 data symbols: ssym(0:7,69).  Re-order the
    bins to remove Gray code.
10. Compute soft symbols for 206 bits (bit 207 is always 0).
11. Remove interleaving
12. Pack bits into bytes, send to Fano decoder
13. If Fano succeeds, remove source encoding and display user message.