fix(calls): wire DTLN ML denoise correctly via @workadventure JS API

The prior DTLN attempt (89a2321d) broke the build (missing dep, wrong
`cinny/` asset paths) and typecheck (`'dtln'` not in DenoiseModelId), and was
wired against an API the package doesn't expose. @workadventure/noise-
suppression is not a flat AudioWorklet — it's a self-contained ES module whose
processor name is `workadventure-noise-suppression` and which resolves its own
LiteRT WASM + TFLite models via import.meta.url. Driving it by hand-rolled
addModule + processorOptions cannot work.

- Re-add @workadventure/noise-suppression@0.0.4 (package.json + lockfile).
- vite: copy the package's whole dist/ tree intact to
  denoise/workadventure/ (preserving assets/ + vendor/litert) so import.meta
  resolution works at runtime; fail the build if the entry module is missing.
- shim: for the DTLN model, dynamic-import denoise/workadventure/audio-worklet
  .js and use createNoiseSuppressionAudioWorklet(ctx, { bypassUntilReady })
  to build the node; RNNoise/Speex keep their direct flat-worklet path. Async
  init errors are logged + reported and fall back to the raw mic.
- Restore 'dtln' in DenoiseModelId (+ settings coercion), the model chart, and
  the settings dropdown, labelled "(beta)".

DTLN builds and is fully self-hosted, but its in-call audio is UNVERIFIED in
this environment — needs a real-call test. DeepFilterNet stays excluded (CDN
asset loading, incompatible with self-hosting / Tauri CSP).

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

build/lotus-denoise.js

@@ -54,6 +54,13 @@
       script: 'speexWorklet.js',
       wasm: 'speex.wasm',
     },
+    dtln: {
+      // @workadventure/noise-suppression is a self-contained ES module that
+      // resolves its own AudioWorklet processor + LiteRT WASM + TFLite models
+      // via import.meta.url. We dynamic-import this helper and let it build the
+      // node, rather than addModule-ing a flat worklet ourselves.
+      helper: 'workadventure/audio-worklet.js',
+    },
     gate: {
       name: '@sapphi-red/web-noise-suppressor/noise-gate',
       script: 'noiseGateWorklet.js',
@@ -113,8 +120,10 @@
           } catch (e) {}
           return Promise.reject(new Error('SampleRate mismatch: ' + ctx.sampleRate));
         }
-        // Load required modules
-        var scripts = [PROCESSORS[MODEL].script];
+        // Load worklet modules. DTLN registers its own processor via the
+        // dynamic-imported helper (see buildMlNode), so it needs nothing here.
+        var scripts = [];
+        if (MODEL === 'rnnoise' || MODEL === 'speex') scripts.push(PROCESSORS[MODEL].script);
         if (USE_GATE) scripts.push(PROCESSORS.gate.script);
 
         return Promise.all(
@@ -138,6 +147,36 @@
 
   var hasNotifiedActive = false;
 
+  // Build the ML denoise AudioWorkletNode. RNNoise/Speex are flat sapphi
+  // worklets we instantiate directly with the fetched WASM binary. DTLN comes
+  // from @workadventure's self-contained helper, which we dynamic-import; it
+  // resolves its own processor + LiteRT WASM + TFLite models internally and
+  // returns the node. Resolves to { node, ready, dispose }.
+  function buildMlNode(ctx, wasmBinary) {
+    if (MODEL === 'dtln') {
+      return import(ASSET_BASE + PROCESSORS.dtln.helper).then(function (mod) {
+        // bypassUntilReady: pass raw audio through until the model is loaded so
+        // the call never has a silent/missing track during init.
+        return mod.createNoiseSuppressionAudioWorklet(ctx, { bypassUntilReady: true });
+      });
+    }
+    var node = new AudioWorkletNode(ctx, PROCESSORS[MODEL].name, {
+      channelCount: 1,
+      numberOfInputs: 1,
+      numberOfOutputs: 1,
+      processorOptions: { maxChannels: 1, wasmBinary: wasmBinary },
+    });
+    return Promise.resolve({
+      node: node,
+      ready: Promise.resolve(),
+      dispose: function () {
+        try {
+          node.port.postMessage('destroy');
+        } catch (e) {}
+      },
+    });
+  }
+
   function processStream(stream) {
     var audioTracks = stream.getAudioTracks();
     if (audioTracks.length === 0) return Promise.resolve(stream);
@@ -166,68 +205,72 @@
         }
 
         // 2. ML Processor
-        var mlNode = new AudioWorkletNode(ctx, PROCESSORS[MODEL].name, {
-          channelCount: 1,
-          channelCountMode: 'explicit',
-          numberOfInputs: 1,
-          numberOfOutputs: 1,
-          outputChannelCount: [1],
-          processorOptions: { maxChannels: 1, wasmBinary: wasmBinary },
-        });
-        head.connect(mlNode);
-        mlNode.connect(dest);
+        return buildMlNode(ctx, wasmBinary).then(function (ml) {
+          var mlNode = ml.node;
+          head.connect(mlNode);
+          mlNode.connect(dest);
 
-        var origTrack = audioTracks[0];
-        var processedTrack = dest.stream.getAudioTracks()[0];
+          // Surface async init failures (e.g. DTLN model load) without blocking
+          // the track handoff — audio flows via bypassUntilReady meanwhile.
+          if (ml.ready && typeof ml.ready.then === 'function') {
+            ml.ready.catch(function (err) {
+              var m = err instanceof Error ? err.message : String(err);
+              console.error('[lotus-denoise] ' + MODEL + ' init failed:', m);
+            });
+          }
 
-        var torndown = false;
-        function cleanup() {
-          if (torndown) return;
-          torndown = true;
-          try {
-            mlNode.port.postMessage('destroy');
-          } catch (e) {}
-          try {
-            source.disconnect();
-            mlNode.disconnect();
-          } catch (e) {}
-          try {
-            origTrack.stop();
-          } catch (e) {}
-        }
+          var origTrack = audioTracks[0];
+          var processedTrack = dest.stream.getAudioTracks()[0];
 
-        var rawStop = processedTrack.stop.bind(processedTrack);
-        processedTrack.stop = function () {
-          cleanup();
-          rawStop();
-        };
-        origTrack.addEventListener('ended', function () {
-          try {
+          var torndown = false;
+          function cleanup() {
+            if (torndown) return;
+            torndown = true;
+            try {
+              ml.dispose();
+            } catch (e) {}
+            try {
+              source.disconnect();
+              mlNode.disconnect();
+            } catch (e) {}
+            try {
+              origTrack.stop();
+            } catch (e) {}
+          }
+
+          var rawStop = processedTrack.stop.bind(processedTrack);
+          processedTrack.stop = function () {
+            cleanup();
             rawStop();
-          } catch (e) {}
-          cleanup();
-        });
+          };
+          origTrack.addEventListener('ended', function () {
+            try {
+              rawStop();
+            } catch (e) {}
+            cleanup();
+          });
 
-        if (!hasNotifiedActive) {
-          hasNotifiedActive = true;
-          window.parent.postMessage(
-            {
-              type: 'lotus-denoise-status',
-              active: true,
-              model: MODEL,
-              nativeNS: USE_NATIVE_NS,
-              gate: USE_GATE,
-            },
-            '*',
-          );
-        }
+          if (!hasNotifiedActive) {
+            hasNotifiedActive = true;
+            window.parent.postMessage(
+              {
+                type: 'lotus-denoise-status',
+                active: true,
+                model: MODEL,
+                nativeNS: USE_NATIVE_NS,
+                gate: USE_GATE,
+              },
+              '*',
+            );
+          }
 
-        var out = new MediaStream();
-        out.addTrack(processedTrack);
-        stream.getVideoTracks().forEach(function (t) {
-          out.addTrack(t);
+          var out = new MediaStream();
+          out.addTrack(processedTrack);
+          stream.getVideoTracks().forEach(function (t) {
+            out.addTrack(t);
+          });
+          return out;
         });
-        return out;
       })
       .catch(function (e) {
         var msg = e instanceof Error ? e.message : String(e);

package-lock.json

@@ -26,6 +26,7 @@
         "@tanstack/react-query-devtools": "5.100.13",
         "@tanstack/react-virtual": "3.13.25",
         "@types/dompurify": "3.2.0",
+        "@workadventure/noise-suppression": "0.0.4",
         "await-to-js": "3.0.0",
         "badwords-list": "2.0.1-4",
         "blurhash": "2.0.5",
@@ -4856,6 +4857,15 @@
         }
       }
     },
+    "node_modules/@workadventure/noise-suppression": {
+      "version": "0.0.4",
+      "resolved": "https://registry.npmjs.org/@workadventure/noise-suppression/-/noise-suppression-0.0.4.tgz",
+      "integrity": "sha512-v8DQgV2TQAWh7YLo7bZ1grV3iDNltRuvPaIYTcaBWoOjUaxDp/j5zrFLz4ZuijPGxzqcQxeW7ql/HJltMuLDtA==",
+      "license": "MIT",
+      "dependencies": {
+        "fft.js": "^4.0.4"
+      }
+    },
     "node_modules/@xobotyi/scrollbar-width": {
       "version": "1.9.5",
       "resolved": "https://registry.npmjs.org/@xobotyi/scrollbar-width/-/scrollbar-width-1.9.5.tgz",
@@ -7619,6 +7629,12 @@
         }
       }
     },
+    "node_modules/fft.js": {
+      "version": "4.0.4",
+      "resolved": "https://registry.npmjs.org/fft.js/-/fft.js-4.0.4.tgz",
+      "integrity": "sha512-f9c00hphOgeQTlDyavwTtu6RiK8AIFjD6+jvXkNkpeQ7rirK3uFWVpalkoS4LAwbdX7mfZ8aoBfFVQX1Re/8aw==",
+      "license": "MIT"
+    },
     "node_modules/file-entry-cache": {
       "version": "8.0.0",
       "resolved": "https://registry.npmjs.org/file-entry-cache/-/file-entry-cache-8.0.0.tgz",

package.json

@@ -50,6 +50,7 @@
     "@tanstack/react-query-devtools": "5.100.13",
     "@tanstack/react-virtual": "3.13.25",
     "@types/dompurify": "3.2.0",
+    "@workadventure/noise-suppression": "0.0.4",
     "await-to-js": "3.0.0",
     "badwords-list": "2.0.1-4",
     "blurhash": "2.0.5",

src/app/features/settings/general/General.tsx

@@ -1441,6 +1441,7 @@ function Calls() {
                   options={[
                     { value: 'rnnoise', label: 'RNNoise' },
                     { value: 'speex', label: 'Speex (Legacy)' },
+                    { value: 'dtln', label: 'DTLN (beta)' },
                   ]}
                 />
               }

src/app/state/settings.ts

@@ -14,10 +14,10 @@ export type MessageSpacing = '0' | '100' | '200' | '300' | '400' | '500';
 //  - 'browser' : WebRTC built-in suppression (Element Call noiseSuppression param)
 //  - 'ml'      : client-side RNNoise ML suppression (Lotus denoise shim)
 export type NoiseSuppressionMode = 'off' | 'browser' | 'ml';
-// Only self-hostable, build-bundled models are exposed. DTLN/DeepFilterNet were
-// evaluated but rely on remote-style asset loading incompatible with our
-// self-hosted/Tauri-CSP strategy (see LOTUS_DENOISE_ENGINEERING_REVIEW.md).
-export type DenoiseModelId = 'rnnoise' | 'speex';
+// Self-hostable, build-bundled ML models. DeepFilterNet remains excluded — it
+// loads its runtime/models from external CDNs, which breaks the self-hosted /
+// Tauri-CSP strategy (see LOTUS_DENOISE_ENGINEERING_REVIEW.md).
+export type DenoiseModelId = 'rnnoise' | 'speex' | 'dtln';
 export type ChatBackground =
   | 'none'
   | 'blueprint'
@@ -263,7 +263,9 @@ export const getSettings = (): Settings => {
       // Coerce any retired/unknown persisted model (e.g. 'dtln', 'deepfilternet'
       // from earlier beta builds) back to the default working model.
       callDenoiseModel:
-        saved.callDenoiseModel === 'rnnoise' || saved.callDenoiseModel === 'speex'
+        saved.callDenoiseModel === 'rnnoise' ||
+        saved.callDenoiseModel === 'speex' ||
+        saved.callDenoiseModel === 'dtln'
           ? saved.callDenoiseModel
           : defaultSettings.callDenoiseModel,
       composerToolbarButtons: {

src/app/utils/lotusDenoiseUtils.ts

@@ -31,6 +31,15 @@ export const DENOISE_MODELS: DenoiseModel[] = [
     transients: 'Poor',
     voiceQuality: 'Moderate',
   },
+  {
+    id: 'dtln',
+    name: 'DTLN (beta)',
+    description: 'Deep-learning model (TFLite). Stronger on transient noise; higher CPU.',
+    cpuUsage: '10-20%',
+    binarySize: '~4 MB',
+    transients: 'Excellent',
+    voiceQuality: 'High',
+  },
 ];
 
 export const isMLDenoiseSupported = (): boolean => {

vite.config.js

@@ -106,6 +106,23 @@ function lotusDenoise() {
       }
       assets.forEach(([s, d]) => fs.copyFileSync(s, d));
 
+      // DTLN (@workadventure/noise-suppression): unlike the flat sapphi
+      // worklets, this package is a self-contained ES module that resolves its
+      // own AudioWorklet processor, LiteRT WASM runtime and TFLite models via
+      // `import.meta.url`. So we copy its whole dist/ tree intact to
+      // denoise/workadventure/ (preserving assets/ + vendor/) and let the shim
+      // dynamic-import denoise/workadventure/audio-worklet.js at runtime. The
+      // entry module is required — fail the build if the install is broken.
+      const dtlnSrc = path.resolve('node_modules/@workadventure/noise-suppression/dist');
+      const dtlnEntry = path.join(dtlnSrc, 'audio-worklet.js');
+      if (!fs.existsSync(dtlnEntry)) {
+        throw new Error(
+          `[lotus-denoise] DTLN entry missing (${dtlnEntry}) — build aborted. ` +
+            'Run `npm ci` to install @workadventure/noise-suppression.',
+        );
+      }
+      fs.cpSync(dtlnSrc, path.join(denoiseDir, 'workadventure'), { recursive: true });
+
       const shimSrc = path.resolve('build/lotus-denoise.js');
       if (!fs.existsSync(shimSrc)) {
         throw new Error(`[lotus-denoise] Missing shim source ${shimSrc} — build aborted.`);