- Model dropdown is now ordered by quality/CPU, best first (DeepFilterNet 3 →
DTLN → RNNoise → Speex); fix RNNoise's inaccurate "High" voice-quality label.
- When a user opts into the ML tier, default to the highest-quality model
(DeepFilterNet 3). The tier default stays browser-native (known-good, best
perceived in testing so far).
- Wire the "Series Suppression" (native-NS-before-ML) toggle into the real call
path — it was applied only in the settings tester, so the tester could sound
better than the actual call. Default it OFF (a single NS stage is best
practice; it's an opt-in test aid).
- isMLDenoiseSupported now also requires WebAssembly, so ML isn't offered on
strict-CSP shells where it would silently fall back to the raw mic.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
- scheduledMessages.test.ts (9): pins the MSC4140 request shape (PUT to the room
send endpoint with the org.matrix.msc4140.delay query, POST cancel/restart to
/delayed_events with the unstable prefix), the delay-floor math (Math.max(1000,
round(sendAt-now)) — "now"/past targets still yield a valid >=1000ms delay),
rounding, and url-encoding.
- lotusDenoiseUtils.test.ts (9): model-catalog data integrity + isMLDenoiseSupported
feature detection across AudioContext/webkit/getUserMedia.
- Bug found + fixed: isMLDenoiseSupported used `!!AudioWorkletNode`, a bare global
reference that throws ReferenceError (not returns false) on a browser with
AudioContext but no AudioWorkletNode binding. Switched to `typeof` so the
detection helper reports unsupported instead of throwing. Regression test proven
to fail on the old code.
Suite now 545 tests (4th real bug caught by the prevention work).
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>