diff --git a/tts_bridge.js b/tts_bridge.js
index 22db2c3..bfaadd6 100644
--- a/tts_bridge.js
+++ b/tts_bridge.js
@@ -189,9 +189,13 @@ function spellWord(word) {
 }
 
 // ── English punctuation → Japanese pause equivalents ──────────────────
+// Characters AquesTalk1 can safely pronounce (kana + limited punctuation).
+// Anything else (！「」・　 etc.) is collapsed to a 、 pause in synthesize().
+const AQTK_SAFE_RE = /[\u3040-\u309F\u30A0-\u30FF\uFF65-\uFF9F\u3001\u3002\uFF1F\u301C\u30FC\u309B\u309C]/;
+
 const PUNCT_MAP = {
-  ",": "、", ".": "。", "!": "！", "?": "？",
-  "，": "、", "。": "。", "！": "！", "？": "？",
+  ",": "、", ".": "。", "!": "、", "?": "？",
+  "，": "、", "。": "。", "！": "、", "？": "？",
   " ": "、", // AquesTalk truncates on space; use 、 pause instead
 };
 
@@ -269,6 +273,14 @@ function synthesize(kanaText) {
   // Collapse consecutive pauses; strip any residual spaces (AquesTalk truncates on space)
   result = result.replace(/ +/g, "、").replace(/、+/g, "、").replace(/^、|、$/g, "");
 
+  // Replace any remaining chars AquesTalk1 cannot pronounce with a 、 pause
+  // (covers ！「」・　 etc. that slipped through from user templates)
+  let cleaned = "";
+  for (const ch of result) {
+    cleaned += AQTK_SAFE_RE.test(ch) ? ch : "、";
+  }
+  result = cleaned.replace(/、+/g, "、").replace(/^、|、$/g, "");
+
   console.error(`[bridge] SYNTH IN:  ${trimmed}`);
   console.error(`[bridge] SYNTH OUT: ${result}`);