Binary FSK: the naval broadcast (STANAG 4481) and MIL-STD-188-110B 5.1

hfmodem.bateleur.fsk sends and receives binary FSK as STANAG 4481 (Edition 1) annex C gives it for NATO naval shore-to-ship broadcasts, with the characteristic frequencies and conventions of MIL-STD-188-110B 5.1 and 4.2. It was written from those texts. No other implementation was read, and no recording of a broadcast has been tested against it.

What the texts give

What neither says, and what is chosen here

• The keying rate of the FSK broadcast. 75 Bd, the rate 4481 annex D keys its FSK test interferer at, is the one tested most here; every rate in the list works.
• Bytes. A broadcast carries a synchronous stream, with no framing in either text. By default a message is sent as start-stop characters: a space start bit, eight data bits, and a mark stop bit, as teleprinters frame them, so that the far end finds the bytes. An S after the mode (75WS) sends the bare stream instead. A receiver joining a stream cannot tell where its bytes begin, so it starts at the first space after the idle line.
• Around a message, half a second of mark (the idle line) goes ahead and two characters' worth follow. The keying is phase-continuous.

On the station, a mode is keyed only if its tones and keying sidebands lie in the span the station declares for bateleur (365–2985 Hz). So multi-channel sub-carriers 1, 15 and 16, and any pair twinned with them, are for offline work (offair) only.

The receiver

stream.ReceiverFsk streams, and fsk.receive runs it over a recording. The mode is set beforehand at both ends: nothing on the air says it.

• Finding it. Every quarter second it looks at the last half second. Where the tones stand in its spectrum gives the carrier offset, within a quarter of the shift (at most 60 Hz) either way. Whether they stand over the places beside them, where nothing is sent, says if a signal is there.
• Soft bits. At every sample, each tone's energy over the bit-long window from there. The soft bit is the mark's share less the space's; twinned channels' are added. The noise reference is averaged over half a second.
• Timing. Where a window lies across a bit's edge the tones mix and their difference shrinks; a bit is read half a bit from where it shrinks most, over the last 96 bits, weighted by the tones' energy so that a stretch where one tone has faded counts for little. On the idle line there are no edges and the timing goes where it likes; once edges are seen it moves a little at a time. Bits are decided 16 bits behind what has arrived.
• Characters. Nothing is taken until eight marks in a row, the idle line, have been read: the bits read as the signal comes up are not bits. After a stop bit that is not mark, characters sent back to back would let a wrong framing pass by chance. So the receiver hunts: it takes the framing under which the next six characters all start with a space and stop with a mark.
• The end. The signal has gone when no bit has stood over the noise for a second; a fade does not end it. Bits are released once a later one has stood clearly over the noise, halfway in decibels between the noise and the signal's own level, so the noise after the end does not become characters.

How it is tested

Against itself only, through bateleur.sim's channels, scored by character error rate (edit distance over the message):

Three runs each, of a 159-character message:

The poor channel's two paths, 2 ms apart, fade the mark and space tones independently, and nothing here is coded. Even read at the true timing, 75W there makes about 0.3% bit errors at 20 dB, which is several percent of ten-bit characters. Diversity (V…D) is the broadcast's own answer to that. Two minutes of noise start nothing. Fed in odd-sized pieces, back-to-back messages come out whole. On the station, a message from the host port keys through the arbiter, split across key-ups at 75 Bd, and decodes back; a replayed transmission reaches the host port.