Tentative question......digital summing
There seem to be quite a few people who frequent this forum who come across very knowledgeable in terms of DSP. I'll start by saying I don't want this to descend into the netherworld of pointless argument, as seems to happen so often when this question has been raised on other forums.
Does anyone know how much truth there could be in the belief that certain audio software "sounds different" than others when only summing is taking place? By this I mean disregarding pan-law differences, having all gains at unity, no plugin processing and DAWs which use PCI/hardware processing (ala PT HD).
I'm assuming that the calculations involved may not be quite as simple as just addition of streams of binary (maybe it is?) but even so, if different results can be achieved with the same input, that equates to error right? Or, is the whole process intensive with high track-counts and therefore different optimization techniques cause slightly differing results......or does it have something to do with round-off error in 32-bit calculations?
It would be useful if someone could point me in the direction of some books/papers that so I could learn more myself also!
Like I said, I don't want the thread to get subjective, any informative replies about the technical aspects would be great!
Cheers for reading all the questions!
- Tim
There is almost certainly some truth to be found here.
Samples can be represented many different ways, e.g. floating point of various word lengths, 16/24 bit PCM, etc.
This can take quite a bit of headroom, which is pretty easy to do in a floating point representation, but somewhat tricky in a straight PCM world.
At the end of the summing process, the result has to be converted back to some representation that makes sense for the intended purpose, e.g. 16 bit, 44.1 KHz for CD.
For best results, Dithering will probably want to be applied whenever an operation that changes the bit width occurs.
Things that might realistically make a difference to the calculation of audio in audio software, and so *theoretically* could be the cause of perceived differences in sound:
1 - calculation bit depth and choice of fixed or floating point arithmetic
2 - floating point round off error and in relation to this order of operation - floating point maths is quite susceptible to different orders of operation. Whilst this is a fact I am aware of - I've not personally to date had enough reason to take this into account when designing algorithms. When for instance you add floating point numbers that are quite different in magnitude from one another the error may become more of a problem (depending on the scenario) - so, one would expect a summing routine that has a binary tree structure (add 2, add another 2 - sum those to get 4 - add to another similarly produced group of 4 etc. etc.) to give different results to a cumulating buffer approach (where we add each new layer to the result of the previously summed layer)s. There may be someone here who can comment further on this....
3 - interpolation types (may only apply if the sample rate is being changed / varispeed playback - but it could be an issue - it is certainly possible to hear the difference between certain kinds of interpolation)
4 - as chris says - final bit depth of output and dithering choices etc.
A.