Tatsh / smlab / 32616031394

100%

push

github

warp: put beat zero where the attacks are, not where the envelope averages out

The grid origin came from the argument of the envelope's Fourier coefficient at
the beat frequency. That is quick and it is a weighted average over everything
playing, so a long decay or a busy off-beat pulls it away from the attacks and it
lands late. It now takes the phase that gathers the most onset strength onto the
grid, then averages the attacks that phase caught to settle where in the window
it belongs, because every phase within the window scores about the same and the
best of a flat top can sit at either edge of it.

Measured against a track whose beat is at 3.3618 s: the gathering score puts that
phase at 1.000 and a phase a third of a beat away at 0.688, and the refined
answer lands within five milliseconds. The old estimate on the same track was out
by a third of a beat.

No change to how --offset behaves. Checking that took some doing and it was right
all along: rendered at an offset of 3.700 s, the bar line is drawn at 3.700 s to
within two milliseconds, and the waveform beside it correlates with the audio at
a shift of one pixel column.

Signed-off-by: Andrew Udvare <audvare@gmail.com>

766 of 798 branches covered (95.99%)

Branch coverage included in aggregate %.

1 of 11 new or added lines in 1 file covered. (9.09%)

3746 of 3832 relevant lines covered (97.76%)

2.93 hits per line