The terms 'Stream separation' or 'Stream segregation" are used here,
meaning more or less the same as polyphonic voice separation or voice
following.

The core algorithm is based on ML clustering techniques, using 'single
link agglomerative clustering' to group individual notes into streams
of voices.

Any overlapping notes are clustered into separate streams, one note
belonging to one stream only.  A dynamic variable *overlap-tolerance*
holds a factor (0-1.0) to control how forgiving the algorithm is with
respect to what counts as simultaneous.

The feature vectors compared here are built from pitch and time info
in input data, using a function to measure vector distances.  The
default distance measure is a euclidian metric.

The default distance function - #'euclidian-distance - compares
pitches using a mel frequency scale, to measure distances closer to
what's considered perceptually significant, and time represented in
milliseconds.  Adjustable weightings for pitch and time allow tuning
of distance measurements.

Typical input is a chord-seq structure.  Output a list of chord-seq's,
one per voice/stream found.

cpu time and memory consumption rises exponentially with larger input
data sets.  If this becomes a problem select subsets of data to work
on, or use the provided analysis-class: 'stream-seg' to interactively
segment data before applying the stream-separation analysis per
segment.

A sub-class of 'analysis' - 'stream-seg' - is provided to work
interactively with stream-segregation within segmentation bounds of
segmented data.  Functions are provided to return the output from this
analysis as a list of lists of chord-seqs, or the whole sequency of
analysed segments concatenated into one multi-seq.

Anders Vinjar - April 2017