How far ONE PAIR OF WINDOWS carries evidence: the delay at which the analysis window's own
overlap drives coherence down to ALIGN_TRUST_COHERENCE, in ms.
MEASURED, and it is a different number from the reach above — the corpus found it the way §9b
found the wrap. Two Hann-windowed blocks D samples apart share only n - D samples of
signal; the rest of each block is content the other never saw. So the coherence a perfect
measurement of a perfect room can reach is capped by the window's own autocorrelation,
rho(d) below with d = D / n, and it falls from 1 at zero delay to 0 at a whole window —
~0.99 at 1 m, ~0.68 at 10 m, ~0.08 at 25 m. It crosses the trust floor at d = 0.228:
~38.9 ms, 13.3 m, again at every graph rate.
Nothing here is a threshold anyone chose. rho is the closed form of a Hann window's
autocorrelation and the floor is the one the desk already declares, so this is those two
facts multiplied out.
IT IS NOT THE FEATURE'S REACH — it is the reach of one pair of windows read at one instant,
and speakerAlignAim is what stops that being the same thing. Reading the reference
into its own past restores the overlap, so this number becomes the SEARCH STEP: the furthest
an aim may sit from the truth and still be believed. The reach is the whole window
(speakerAlignReachMs).
How far ONE PAIR OF WINDOWS carries evidence: the delay at which the analysis window's own overlap drives coherence down to ALIGN_TRUST_COHERENCE, in ms.
MEASURED, and it is a different number from the reach above — the corpus found it the way §9b found the wrap. Two Hann-windowed blocks
Dsamples apart share onlyn - Dsamples of signal; the rest of each block is content the other never saw. So the coherence a perfect measurement of a perfect room can reach is capped by the window's own autocorrelation,rho(d)below withd = D / n, and it falls from 1 at zero delay to 0 at a whole window — ~0.99 at 1 m, ~0.68 at 10 m, ~0.08 at 25 m. It crosses the trust floor atd = 0.228: ~38.9 ms, 13.3 m, again at every graph rate.Nothing here is a threshold anyone chose.
rhois the closed form of a Hann window's autocorrelation and the floor is the one the desk already declares, so this is those two facts multiplied out.IT IS NOT THE FEATURE'S REACH — it is the reach of one pair of windows read at one instant, and speakerAlignAim is what stops that being the same thing. Reading the reference into its own past restores the overlap, so this number becomes the SEARCH STEP: the furthest an aim may sit from the truth and still be believed. The reach is the whole window (speakerAlignReachMs).