Deploy Ekko v1 Tiny + PnC browser demo

app.js

@@ -92,7 +92,7 @@ const state = {
 };
 
 const worker = new Worker(signedAsset("worker.js?v=20260927-2"));
-const pncWorker = new Worker(signedAsset("pnc-worker.js?v=20260927-2"));
+const pncWorker = new Worker(signedAsset("pnc-worker.js?v=20261001-1"));
 
 function setStatus(text, detail = "", kind = "loading") {
   elements.statusText.textContent = text;

pnc-worker.js

@@ -21,8 +21,8 @@ let initialization = null;
 let session = null;
 let vocabulary = null;
 let labels = null;
-let maximumLength = 256;
-const INFERENCE_WINDOW_TOKENS = 192;
+let maximumLength = 128;
+const INFERENCE_WINDOW_TOKENS = 128;
 let pendingRequest = null;
 let processing = false;
 
@@ -101,7 +101,7 @@ async function initialize(configUrl) {
   vocabulary = tokenizer.model.vocab;
   labels = modelConfig.id2label;
   // Shorter context restores sentence stops in long, continuous dictation.
-  maximumLength = Math.min(config.maxLength || 256, INFERENCE_WINDOW_TOKENS);
+  maximumLength = Math.min(config.maxLength || 128, INFERENCE_WINDOW_TOKENS);
   session = await ort.InferenceSession.create(downloaded.get("pnc.int8.onnx"), {
     executionProviders: ["wasm"],
     graphOptimizationLevel: "all",
@@ -172,16 +172,26 @@ function encodeWords(words) {
   return {identifiers, wordIds};
 }
 
-function encodedLength(words) {
-  return encodeWords(words).identifiers.length;
+function windowTokenIndex(words) {
+  const {wordIds} = encodeWords(words);
+  const counts = new Array(words.length).fill(0);
+  let specialTokens = 0;
+  for (const wordId of wordIds) {
+    if (wordId === null) specialTokens += 1;
+    else counts[wordId] += 1;
+  }
+  const prefix = [0];
+  for (const count of counts) prefix.push(prefix.at(-1) + count);
+  return {prefix, specialTokens};
 }
 
-function maximumWindowEnd(words, start) {
+function maximumWindowEnd(words, start, prefix, specialTokens) {
+  const tokenBudget = prefix[start] + maximumLength - specialTokens;
   let low = start + 1;
   let high = Math.min(words.length, start + maximumLength);
   while (low < high) {
     const middle = Math.floor((low + high + 1) / 2);
-    if (encodedLength(words.slice(start, middle)) <= maximumLength) {
+    if (prefix[middle] <= tokenBudget) {
       low = middle;
     } else {
       high = middle - 1;
@@ -253,15 +263,16 @@ async function predictWindow(words) {
 
 async function punctuateWords(words) {
   if (!words.length) return [];
+  const {prefix, specialTokens} = windowTokenIndex(words);
   const output = [];
   let nextWord = 0;
-  const overlapWords = 24;
+  const overlapWords = 12;
   while (nextWord < words.length) {
     let contextStart = Math.max(0, nextWord - overlapWords);
-    let windowEnd = maximumWindowEnd(words, contextStart);
+    let windowEnd = maximumWindowEnd(words, contextStart, prefix, specialTokens);
     if (windowEnd <= nextWord) {
       contextStart = nextWord;
-      windowEnd = maximumWindowEnd(words, contextStart);
+      windowEnd = maximumWindowEnd(words, contextStart, prefix, specialTokens);
     }
     const rendered = await predictWindow(words.slice(contextStart, windowEnd));
     const emitEnd = windowEnd === words.length