Design.txt

Hi! Since I've opened up the source for Digits, I wanted to take a second to
talk about some of the design decisions I made while writing it.

If you didn't know, Digits started out as a simple project for me to work on
while I was hanging out at hacker spaces. It started pretty modestly. The
first version only sported one LFO, only had the delay effect, no filters,
and the envelopes were trapezoidal. It has come a long way!

One thing I was particularly happy with was how the phase distortion algorithm
came out. While some examples I've seen use different routines for different
waveshapes, Digits uses a consistent formula and just changes settings to
produce all of its waveforms. For example, saw means you run the shaper
once per cycle. Square means you run the shaper twice per cycle. Skew means
that you alter position of the node where the knee in the phase distortion is.
The pulse-width modulation is faked in by applying a second shaper to the
entire generator (so, a shaper within a shaper). This is also why the
pulse-width modulation causes the waveform to brighten as the pulse narrows.

The design of Digits is not without some regrets: I went with a control rate
for the main audio loop instead of making it 100% sample-accurate. This is
because that's how I always did synthesizers, and this is how a lot of older
synths operated. It makes for a messy loop and causes a tiny amount of jitter.
A much better way of doing it would be to render the number of samples between
MIDI events precisely.

A big problem in the user interface is the Shaper Lock in the filter section.
It provides a lot of flexibility, but at the cost of being unorthodox and
confusing to many users.

My final regret that I want to talk about is how much of a mess the UI code
is! It really grew a lot. If I did it again, I'd leverage classes more to
clean it up. This ended up looking more the way it'd look if you just wrote
it in flat C. Oops.

Optimization: While the oscillators look like they are doing a lot of heavy
lifting, they are actually fairly fast. However, the LFOs and especially the
envelopes are slow. I would suggest making a table-driven envelope to replace
it. Be aware though that the attack is an inverse curve from the decay shape.

You might also notice that the oscillator uses an uninterpolated lookup. I
investigated on a few devices, seeing whether the increased table size might
cause a CPU cache miss, or whether it would drastically hurt the sound
quality. It was very fast without interpolation and did not appear to hurt
the sound quality too much, so that is what I went for. My secret shame.

Lastly, there are many, many areas where you could improve the speed with
vector processing. However, this particular version of the Digits engine only
uses basic C++ code and no inline assembly or vector intrinsics.

As far as sound quality goes, it can alias a little bit at 44 or 48khz. I
started on an oversampling routine but never fit it into the code (this is
all done as an after-work hobby after all). Though aliasing is combated a bit
by toning down the shaper at higher pitches, a little oversampling might
go a long way.

Anyway, that's about it! Happy hacking; be sure to read the License.txt!
(Digits is licensed under GPL)

Cheers,
Louis "Extent of the Jam" Gorenfeld
louis.gorenfeld@gmail.com