@utterance/core 0.0.2 → 0.0.4

package/dist/index.cjs

@@ -373,7 +373,7 @@ var FeatureExtractor = class {
 };
 
 // src/types.ts
-var MODEL_CDN_URL = "https://pub-46a5feb0029246bcbc93fab6162cff94.r2.dev/v0.0.2/utterance-v1.onnx";
+var MODEL_CDN_URL = "https://pub-46a5feb0029246bcbc93fab6162cff94.r2.dev/v2/utterance-v2.onnx";
 var DEFAULT_OPTIONS = {
   sensitivity: 0.5,
   pauseTolerance: 1500,
@@ -484,7 +484,9 @@ var ONNXModel = class {
         }
       } else if (path === "bundled") {
         try {
-          const response = await fetch(new URL("../../models/utterance-v1.onnx", import_meta.url).href);
+          const getUrl = new Function("p", "b", "return new URL(p, b).href");
+          const href = getUrl("../../models/utterance-v2.onnx", import_meta.url);
+          const response = await fetch(href);
           if (response.ok) {
             modelSource = await response.arrayBuffer();
           } else {
@@ -569,6 +571,25 @@ var ONNXModel = class {
       const dstIdx = i * FEATURE_DIM;
       input.set(this.frameBuffer.subarray(srcIdx, srcIdx + FEATURE_DIM), dstIdx);
     }
+    for (let f = 0; f < 14; f++) {
+      let sum = 0;
+      for (let i = 0; i < CONTEXT_FRAMES; i++) {
+        sum += input[i * FEATURE_DIM + f];
+      }
+      const mean = sum / CONTEXT_FRAMES;
+      let varSum = 0;
+      for (let i = 0; i < CONTEXT_FRAMES; i++) {
+        const d = input[i * FEATURE_DIM + f] - mean;
+        varSum += d * d;
+      }
+      const std = Math.sqrt(varSum / CONTEXT_FRAMES) || 1;
+      for (let i = 0; i < CONTEXT_FRAMES; i++) {
+        input[i * FEATURE_DIM + f] = (input[i * FEATURE_DIM + f] - mean) / std;
+      }
+    }
+    for (let i = 0; i < CONTEXT_FRAMES; i++) {
+      input[i * FEATURE_DIM + 14] /= 500;
+    }
     const tensor = new ort.Tensor("float32", input, [1, CONTEXT_FRAMES, FEATURE_DIM]);
     const results = await session.run({ input: tensor });
     const output = results.output;
@@ -616,6 +637,7 @@ var TurnDetector = class {
   state = "idle";
   pauseStart = 0;
   speakStart = 0;
+  interruptFired = false;
   sensitivity;
   pauseTolerance;
   constructor(sensitivity = 0.5, pauseTolerance = 1500) {
@@ -645,6 +667,7 @@ var TurnDetector = class {
     const threshold = this.sensitivity;
     switch (label) {
       case "speaking":
+        this.interruptFired = false;
         if (this.state !== "speaking") {
           this.state = "speaking";
           this.speakStart = timestamp;
@@ -680,7 +703,8 @@ var TurnDetector = class {
         }
         break;
       case "interrupt_intent":
-        if (confidence >= threshold) {
+        if (confidence >= threshold && !this.interruptFired) {
+          this.interruptFired = true;
           this.emit("interrupt", { timestamp });
         }
         break;
@@ -693,6 +717,7 @@ var TurnDetector = class {
     this.state = "idle";
     this.pauseStart = 0;
     this.speakStart = 0;
+    this.interruptFired = false;
   }
   emit(event, payload) {
     this.listeners.get(event)?.forEach((fn) => fn(payload));

package/dist/index.js

@@ -337,7 +337,7 @@ var FeatureExtractor = class {
 };
 
 // src/types.ts
-var MODEL_CDN_URL = "https://pub-46a5feb0029246bcbc93fab6162cff94.r2.dev/v0.0.2/utterance-v1.onnx";
+var MODEL_CDN_URL = "https://pub-46a5feb0029246bcbc93fab6162cff94.r2.dev/v2/utterance-v2.onnx";
 var DEFAULT_OPTIONS = {
   sensitivity: 0.5,
   pauseTolerance: 1500,
@@ -447,7 +447,9 @@ var ONNXModel = class {
         }
       } else if (path === "bundled") {
         try {
-          const response = await fetch(new URL("../../models/utterance-v1.onnx", import.meta.url).href);
+          const getUrl = new Function("p", "b", "return new URL(p, b).href");
+          const href = getUrl("../../models/utterance-v2.onnx", import.meta.url);
+          const response = await fetch(href);
           if (response.ok) {
             modelSource = await response.arrayBuffer();
           } else {
@@ -532,6 +534,25 @@ var ONNXModel = class {
       const dstIdx = i * FEATURE_DIM;
       input.set(this.frameBuffer.subarray(srcIdx, srcIdx + FEATURE_DIM), dstIdx);
     }
+    for (let f = 0; f < 14; f++) {
+      let sum = 0;
+      for (let i = 0; i < CONTEXT_FRAMES; i++) {
+        sum += input[i * FEATURE_DIM + f];
+      }
+      const mean = sum / CONTEXT_FRAMES;
+      let varSum = 0;
+      for (let i = 0; i < CONTEXT_FRAMES; i++) {
+        const d = input[i * FEATURE_DIM + f] - mean;
+        varSum += d * d;
+      }
+      const std = Math.sqrt(varSum / CONTEXT_FRAMES) || 1;
+      for (let i = 0; i < CONTEXT_FRAMES; i++) {
+        input[i * FEATURE_DIM + f] = (input[i * FEATURE_DIM + f] - mean) / std;
+      }
+    }
+    for (let i = 0; i < CONTEXT_FRAMES; i++) {
+      input[i * FEATURE_DIM + 14] /= 500;
+    }
     const tensor = new ort.Tensor("float32", input, [1, CONTEXT_FRAMES, FEATURE_DIM]);
     const results = await session.run({ input: tensor });
     const output = results.output;
@@ -579,6 +600,7 @@ var TurnDetector = class {
   state = "idle";
   pauseStart = 0;
   speakStart = 0;
+  interruptFired = false;
   sensitivity;
   pauseTolerance;
   constructor(sensitivity = 0.5, pauseTolerance = 1500) {
@@ -608,6 +630,7 @@ var TurnDetector = class {
     const threshold = this.sensitivity;
     switch (label) {
       case "speaking":
+        this.interruptFired = false;
         if (this.state !== "speaking") {
           this.state = "speaking";
           this.speakStart = timestamp;
@@ -643,7 +666,8 @@ var TurnDetector = class {
         }
         break;
       case "interrupt_intent":
-        if (confidence >= threshold) {
+        if (confidence >= threshold && !this.interruptFired) {
+          this.interruptFired = true;
           this.emit("interrupt", { timestamp });
         }
         break;
@@ -656,6 +680,7 @@ var TurnDetector = class {
     this.state = "idle";
     this.pauseStart = 0;
     this.speakStart = 0;
+    this.interruptFired = false;
   }
   emit(event, payload) {
     this.listeners.get(event)?.forEach((fn) => fn(payload));

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@utterance/core",
-  "version": "0.0.2",
+  "version": "0.0.4",
   "description": "Client-side semantic endpointing. Know when they're done talking.",
   "type": "module",
   "main": "dist/index.cjs",
@@ -56,6 +56,8 @@
   "devDependencies": {
     "@eslint/js": "^9.0.0",
     "@tailwindcss/postcss": "^4.2.0",
+    "@types/d3-scale": "^4.0.9",
+    "@types/d3-shape": "^3.1.8",
     "@types/mdx": "^2.0.13",
     "@types/node": "^22.0.0",
     "@types/react": "^19.2.14",
@@ -76,9 +78,11 @@
   "dependencies": {
     "@next/third-parties": "^16.1.6",
     "@react-three/fiber": "^9.5.0",
-    "@utterance/core": "^0.0.1",
+    "@utterance/core": "^0.0.2",
     "class-variance-authority": "^0.7.1",
     "clsx": "^2.1.1",
+    "d3-scale": "^4.0.2",
+    "d3-shape": "^3.2.0",
     "fumadocs-core": "^16.6.3",
     "fumadocs-mdx": "^14.2.7",
     "fumadocs-ui": "^16.6.3",