feat(lotus-denoise): quality — dry/wet attenuation floor, gate-after-ML, softer DFN

Track-B audio-quality changes to reduce the "robotic/underwater" artifact.

- Dry/wet attenuation floor (default 0.15 ≈ -16 dB) blends a little of the raw
  mic under the denoised signal so suppression can't fully collapse the noise
  floor between words (the main cause of the RNNoise "underwater"/pumping
  sound). Applied ONLY to the low-latency flat models (RNNoise/Speex); DTLN/DFN
  add algorithmic latency that would comb-filter an undelayed dry mix, so they
  rely on their own level instead. Tunable via `lotusDenoiseFloor`.
- Noise gate now runs AFTER the ML model, not before — gating the raw signal
  fed hard-zeroed frames into the model and tuned the threshold on pre-denoise
  levels.
- DeepFilterNet 3 noiseReductionLevel 80 -> 60: full strength was the main
  "over-processed" contributor; 60 keeps voice natural.

Defaults are conservative and tunable; final values are meant to be dialed in
with real-call A/B listening.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

src/lotus/lotusDenoise.ts

@@ -73,6 +73,7 @@ export function startLotusDenoise(vm: CallViewModel): () => void {
       : "rnnoise";
 
   const rawThreshold = lotusParam("lotusGateThreshold");
+  const rawFloor = lotusParam("lotusDenoiseFloor");
   const config: LotusDenoiseConfig = {
     model,
     assetBase: safeAssetBase(lotusParam("lotusDenoiseBase")),
@@ -83,6 +84,13 @@ export function startLotusDenoise(vm: CallViewModel): () => void {
       rawThreshold !== null && Number.isFinite(Number(rawThreshold))
         ? Number(rawThreshold)
         : -45,
+    // Dry/wet attenuation floor. Default 0.15 (~-16 dB) tames the
+    // over-suppression "underwater"/pumping artifact; host can tune via
+    // `lotusDenoiseFloor` (0 = full suppression, no floor).
+    floor:
+      rawFloor !== null && Number.isFinite(Number(rawFloor))
+        ? Math.min(0.5, Math.max(0, Number(rawFloor)))
+        : 0.15,
   };
 
   const micOf = (room: LivekitRoom): LocalAudioTrack | undefined =>

src/lotus/lotusDenoiseProcessor.ts

@@ -24,6 +24,13 @@ export interface LotusDenoiseConfig {
   assetBase: string;
   gate: boolean;
   gateThreshold: number;
+  /**
+   * Attenuation floor as a dry/wet mix: the fraction (0..~0.3) of the ORIGINAL
+   * mic blended back under the denoised signal so full suppression never fully
+   * collapses the noise floor — this is what kills the "underwater"/pumping
+   * artifact. 0.15 ≈ a -16 dB floor. 0 = full suppression (no floor).
+   */
+  floor: number;
 }
 
 // Flat sapphi worklets (RNNoise/Speex): each registers a processor under `name`
@@ -241,8 +248,17 @@ export class LotusDenoiseProcessor
     const dest = ctx.createMediaStreamDestination();
     const nodes: AudioNode[] = [];
     const disposes: (() => void)[] = [];
-    let head: AudioNode = source;
 
+    // Wet (denoised) path: source → ml → [gate] → wetGain.
+    const ml = await this.buildMlNode(ctx);
+    source.connect(ml.node);
+    nodes.push(ml.node);
+    if (ml.dispose) disposes.push(ml.dispose);
+    let wetHead: AudioNode = ml.node;
+
+    // Gate AFTER the ML model, not before: gating the raw noisy signal fed hard
+    // zeroed frames into the model (discontinuities it must fight) and made the
+    // threshold operate on pre-denoise levels. Gate the residual instead.
     if (this.config.gate) {
       const gate = new AudioWorkletNode(ctx, GATE.name, {
         processorOptions: {
@@ -252,18 +268,43 @@ export class LotusDenoiseProcessor
           maxChannels: 1,
         },
       });
-      head.connect(gate);
-      head = gate;
+      wetHead.connect(gate);
+      wetHead = gate;
       nodes.push(gate);
     }
 
-    const ml = await this.buildMlNode(ctx);
-    head.connect(ml.node);
-    ml.node.connect(dest);
-    nodes.push(ml.node);
-    if (ml.dispose) disposes.push(ml.dispose);
+    // Only mix a dry floor for the LOW-LATENCY flat models (RNNoise/Speex).
+    // DTLN/DeepFilterNet add tens of ms of algorithmic latency, so summing an
+    // undelayed dry copy would comb-filter the voice — for those we rely on the
+    // model's own level (e.g. DFN noiseReductionLevel) instead. RNNoise is also
+    // where the "robotic/underwater" reports come from, so this targets it.
+    const lowLatency =
+      this.config.model === "rnnoise" || this.config.model === "speex";
+    const floor = lowLatency ? Math.min(0.5, Math.max(0, this.config.floor)) : 0;
+    if (floor > 0) {
+      // Dry/wet mix: blend a small amount of the ORIGINAL mic under the
+      // denoised signal so suppression can't fully collapse the noise floor
+      // (kills the "underwater"/pumping artifact). During speech (denoised ≈
+      // original) the two sum back to ~unity; in noise-only gaps the output
+      // floors at `floor` × original instead of digital silence.
+      const wetGain = ctx.createGain();
+      wetGain.gain.value = 1 - floor;
+      wetHead.connect(wetGain);
+      wetGain.connect(dest);
+      nodes.push(wetGain);
 
-    logger.info(`[lotus] denoise processor active (${this.config.model})`);
+      const dryGain = ctx.createGain();
+      dryGain.gain.value = floor;
+      source.connect(dryGain);
+      dryGain.connect(dest);
+      nodes.push(dryGain);
+    } else {
+      wetHead.connect(dest);
+    }
+
+    logger.info(
+      `[lotus] denoise processor active (${this.config.model}, floor=${floor})`,
+    );
     return { source, nodes, disposes, track: dest.stream.getAudioTracks()[0] };
   }
 
@@ -285,7 +326,10 @@ export class LotusDenoiseProcessor
       const mod = await import(/* @vite-ignore */ `${base}deepfilternet/index.esm.js`);
       const core = new mod.DeepFilterNet3Core({
         sampleRate: 48_000,
-        noiseReductionLevel: 80,
+        // 60, not 80: full-strength suppression is the main source of the
+        // "over-processed" character; a lower level keeps voice natural while
+        // the dry/wet floor handles the noise tail.
+        noiseReductionLevel: 60,
         assetConfig: { cdnUrl: dfnBase },
       });
       await core.initialize();