The waveform from a single free reed is more like getting one "puff" of air each time the reed reaches its full excusion. I would probably start with a square-wave and if possible modify the pulse width down (or up) to increase the higher harmonics. Then I'd apply some equalization (mostly low pass) to round things out and to get the "wigglyness" in the waveform.
The effect that you are trying to achieve with the LFO modulated filter is the amplitude modulation that comes from the multiple reeds that sound when a single key is pressed.
This effect could be (approximately) achieved by using the LFO to modulate the amplitude of the signal (approximately because the beat rate of the modulation will go up with pitch if we assume that the reeds are tuned to +/-XXX cents of one another). This has some advantages in terms of simplicty.
Alternately, one could explicitly model the separate reeds, and put a slight tuning difference between them (on the order of a few cents) to get the modulation. This would allow you to directly model the fact that in most harmonium music, there are multiple reeds sounding when a single key is pressed. Registration changes/differences can also be accommodated in this approach by adding additional, octave separated, generators (probably with different filtering characteristics).
To summarize:
- starting with a (pulse width modulated) square wave is probably better.
- part of the sound comes from the mixing of multiple imperfectly tuned reeds sounding at the same time (sometimes even within a given reed bank); this results in amplitude modulation.
- some registrations involve reeds at different octaves (different reed banks) sounding at the same time.
