Sat Jul 28 20:38:19 EDT 2012

S/D notes

- From [1], not all possible code words are generated in the
  constant-input case.

  Can this be turned around and translated to an optimization problem
  where a signal waveform is generated such that noise is maximally
  decorrelated from signal, possibly storing these in tables?

- Could subsampling/supersampling work, i.e. pitching?  What's the
  effect on noise?  If the samples are not short-time correlated, this
  could possibly work.  ( It seems that a lot of applications arise
  when the correlations can be controlled.. )

- Can an incremental decorrelator work?  I.e. starting from a linear
  approximation, is it possible to "push" parts of the noise to higher

- Does it make sense to use 2-dimensional representations,
  i.e. complex signals?  This gives more "room" to push the noise to?

- Decorrelation: "more" decorrelation corresponds to low-pass
  filtering the signal, and probably also moving the noise into the
  lower bands.  Investigate that latter part.

- Can the decorrelator be used with fractional samples?  Actually, it
  is a sigma-delta converter with non-deterministic outputs.

Paper notes from trip:

- Even when there is no correlation between 2 noise signals and their
  respective signals, XOR (multiplication) always introduces noise
  spectrum shifts based on the highest frequencies.  However, in
  practice this might be controlled (i.e. reso lowpass), and not even
  a problem (human hearing drops to 0, doesn't roll of gently).

- Sound synthesis, why the trouble?  There should be no aliasing due
  to nonlinearities + AND gives a "free" nonlinearity.

[1] isbn://0792393090