Weld's allpass sections add group delay between 500 Hz and 4 kHz, where
listeners hear it best. Liski et al. (2021) measured a threshold of about
0.64 ms for a positive group-delay peak on clicks. Weld now reports how much
its delay varies over that range and, above the threshold, suggests
comparing attacks in SUM and SUM DRY.
Common fits from Weld's own tests spread 0.82-1.03 ms, so the line is a
neutral note, not a warning: sustained bass hides much more than a click.
- Fit::group_delay (analytic, per section) and Fit::smear_ms in weld-core.
- Report::smear_ms, filled by the analyzer from the applied fit.
- Tests against the numeric phase slope and the analog 4Q/w0 peak.
- Display only; the audio path does not change.
Weld hears the other bass layer through the sidechain and fits polarity
plus up to four RBJ allpass sections to the measured cross-spectrum, so
the layers sum instead of cancelling. No latency; the layer on its own
keeps its sound.
- Worker thread: Welch cross-spectrum and coherence, Nelder-Mead fit with
1/f weighting, a per-band loss penalty and model selection by share of
the achievable gain; GCC-PHAT lag detection flags a late layer.
- Audio thread: allocation-free chain with a 40 ms warm-up and 40 ms
crossfade per new fit; mono blocks go to the worker through a lock-free
queue.
- Follow / Learn / Hold; the fit is persisted and restored without a
crossfade.
- Editor: per-band loss and recovery against a perfect alignment, phase
before and after, recovered dB, time offset and status.
- Tests for flat magnitude, crossfade, fit targets, Follow hysteresis,
Learn, silence, realtime allocation and the persisted fit in the host.
nvrl