v2.6.1

Fixes and improvements

AGC and noise gate — edge cases

• AGC no longer walks up its gain on non-speech: periodic denoiser residue and repeated knocks can't accumulate into a "speech" run any more, while quiet syllabic voices still adapt.
• Noise gate now derives its threshold from a rolling noise estimate instead of a fixed one, ignores digital-silence frames, and holds back deep attenuation until it has actually seen speech — so first words and quiet microphones survive.
• gateThreshold / gateReduction are applied to the live gate instead of rebuilding it, and the values passed at construction are honoured.

processBuffer() — Speed preset compatibility

• Fixed speed (and quality) failing in buffer processing: the model now runs at its own supported sample rate, with the buffer resampled into it and back, instead of being forced to the buffer's rate.
• Init failures now report through ErrorBus and reject, instead of silently returning an empty buffer.
• Studio Sound state is rebuilt per call, so AGC state no longer leaks between processBuffer() calls.

Upgrading

Nothing to do. If you use the AGC in a noisy room, expect it to settle lower and steadier than in 2.6.0 — that's the fix.