fix: janky captions experience

js/src/features/particles/transcript-overlay.tsx

@@ -1,4 +1,4 @@
-import { useMemo } from "react";
+import { useMemo, useRef } from "react";
 import type { Transcript } from "@/api/types";
 
 type Sentence = Transcript["paragraphs"][number]["sentences"][number];
@@ -41,15 +41,33 @@ export function TranscriptOverlay({
   const activeWord =
     activeWordIndex !== null ? transcript.words[activeWordIndex] : null;
 
-  // Find which chunk contains the active word
+  // Remember the last spoken word so highlights hold during pauses
+  const lastSpokenWordRef = useRef<Word | null>(null);
+  if (activeWord) {
+    lastSpokenWordRef.current = activeWord;
+  }
+  const highlightWord = activeWord ?? lastSpokenWordRef.current;
+
+  const lastChunkRef = useRef<Word[] | null>(null);
+
+  // Find which chunk contains the active word, holding the last one during pauses
   const activeChunk = useMemo(() => {
-    if (!activeWord) return chunks[0] ?? null;
+    if (activeWord) {
-    for (const chunk of chunks) {
+      for (const chunk of chunks) {
-      if (chunk.some((w) => w.start === activeWord.start && w.end === activeWord.end)) {
+        if (chunk.some((w) => w.start === activeWord.start && w.end === activeWord.end)) {
-        return chunk;
+          lastChunkRef.current = chunk;
+          return chunk;
+        }
       }
     }
-    return chunks[0] ?? null;
+    // No active word (speaker pausing) — hold the last chunk
+    if (lastChunkRef.current && chunks.some((c) => c === lastChunkRef.current)) {
+      return lastChunkRef.current;
+    }
+    // Sentence changed, last chunk no longer valid — use first chunk
+    const fallback = chunks[0] ?? null;
+    lastChunkRef.current = fallback;
+    return fallback;
   }, [chunks, activeWord]);
 
   if (!activeSentence || !activeChunk || activeChunk.length === 0) return null;
@@ -62,7 +80,7 @@ export function TranscriptOverlay({
       <p className="rounded-lg px-5 py-3 text-2xl text-center max-w-lg">
         {activeChunk.map((word, i) => {
           const isSpoken =
-            activeWord !== null && word.start <= activeWord.end;
+            highlightWord !== null && word.start <= highlightWord.end;
 
           return (
             <span