Well, as others have cited, just like you can have a third overtone & 5th overtone crystals, which are crystals at 1/3rd, or 1/5th their written value with a LC tuned oscillator just running at 3x or 5x speed, it might be possible to go in the opposite direction. Get a 32.768 Khz crystal, connect it to an LC oscillator tuned at some odd number dividend, the further away, the more difficult to retain lock, and it should work. This means you basically use a 1 transistor LC oscillator tuned to, example 1/9th 32.768k = 3.640888889 Khz. Then use that oscillator as a reference for the next stage 1 transistor oscillator to divide by 9 again = 404.5432099 hz, then /9 again = 44.94924554 hz. That's 3 transistors to go from 32k to 44.949hz. But, you also need 3 ferrite tunable inductors or 3 varicaps which may drift with temperature and time.
A divide by 3 or 5 per stage would be more tolerant. In theory, maybe even a divide by 15 or even bigger may work. At that level, I would breadboard a 1 transistor fundamental 32Khz, with it's output going through a 5pf to 100pf feeding the base of a second 1 transistor LC oscillator tuned to something like 1/101 the frequency and fiddle around with the inductor's tuning just to see if there is any harmonic locking. It will stand out on a scope as you adjust the frequency (ie tune the inductor), there will be zones where the frequency seems to snap to a fixed value. If you have a REAL frequency counter, it's reading will stabilize in these zones while anywhere else, the frequency will be unstable or drift. This would give you a 32Khz / 101 with only 2 transistors... Or try smaller divisions. A smaller divide by 3 or 5 would snap/lock very hard.
This should be bread-board-able with a couple of 2N3904, 1 watch crystal, some caps, resistors and a TOKO 7 style 2mh adjustable inductor. Maybe buffer the output with a third 2N3904 to isolate your scope probe's capacitance, and as a LPF to remove any 32Khz from the source, as tuning to a /101 will be really touchy.