Abstract
A technique for concealing missed processing deadlines ("dropouts"/"xruns") in real-time digital audio output by synthesizing a plausible continuation of the signal, computed speculatively, instead of silence or a glitch. A short-time Fourier analysis runs continuously over the recent signal history; when concealment is needed, each analysis bin is continued independently by extrapolating its phase and amplitude and, for bins containing two sinusoidal partials, a fitted two-pole autoregressive model - generalizing the classical single-rotator phase-vocoder freeze, exact only for one partial per bin. A derived aliasing limit constrains the analysis hop usable for the two-partial model. A cheap closed-form test checks the per-bin recursion's stability before falling back to explicit root-finding and clamping only when needed. The continuation is spliced onto the real signal using one of two seam-repair strategies - a broadband exponential correction anchored to a short linear predictor, or a short linear predictor substituted at the seam - chosen per occurrence from a spectral-similarity measurement, each optimal for a different class of material. A frequency-damping term suppresses a positive-feedback failure mode arising when previously concealed audio is itself re-analyzed. An optional second, short-window instance of the same model runs in parallel and takes over newly-attacked content, because a long window's history remains contaminated by pre-attack content long after a new note begins. Measured on real sustained musical material at high simulated miss rates, each mechanism measurably improved concealment quality, for the content it specifically targets, over omitting it.
Creative Commons License
This work is licensed under a Creative Commons Attribution 4.0 License.
Recommended Citation
Steppeler, Sebastian, "Spectral-continuation concealment of real-time audio deadline misses", Technical Disclosure Commons, ()
https://www.tdcommons.org/dpubs_series/12624