I suspect the problem lies somewhere between 2 buffers - 1 is a circular buffer, the other a memory mapped buffer.
Basically 1 is "spinning" through the buffer wrapping around at the end
The other copies the block directly as if it were fixed.

Or could be fixed size buffer meets buffer with an "eof" symbol (or runlength header).

Note that this could happen just about any level so is very hard to eliminate completely.

You just shouldn't use gravis ultrasound for stream audio stick to using it for demos and intro's and use sb16 for playing 16bit streaming stuff 😀

But, why?, doesn't GUS take away less cpu cycles than SB (in emulation)?

BTW - What a nasty avatar you have here, sir. 😁

Well the gus can more or less run in 2 modes, you can use it as a mod player with 1mb of samples or you can use the dma transfers to fill a little piece of of the 1 mb memory and put a sample on loop to use it like the dma playing of the soundblaster. Soundblaster will be faster than and with less clicks

And it's a nice avatar, tin tin showing bobby whose boss.

I wonder what exactly fast tracker 2 does here.
AFAIK, as other trackers, it uploads samples to GUS memory
and all are mixed by GUS. Clicks then appear with 16-bit,
but only some notes (i.e if they are resampled in 'hardware').
So I think it has nothing to do with DMA (more like interpolation error at the end/start of a sample or buffer?)

Btw. GUS mixing routines can mix 32 channels without a glitch, with SB it gets overloaded at much lower count.

Hmm.. might be a signed vs unsigned integer error.
BTW, when i tried it on my 386+SB many moons ago,
it also had a high shriek & click rate. (Although it was an ESS Audio vs SB)