diff --git a/web-ui/voice-call.js b/web-ui/voice-call.js
index 75a56fb..0415588 100644
--- a/web-ui/voice-call.js
+++ b/web-ui/voice-call.js
@@ -239,6 +239,18 @@
 
       case 'error':
         console.warn('[voice-call] engine error:', msg.message);
+        // The failed request may have been the one holding ttsBusy — without
+        // resetting it here, pumpTtsQueue() early-returns forever and the
+        // call goes silent for the rest of the session.
+        if (!msg.id || msg.id === currentTtsId) {
+          ttsBusy = false;
+          currentTtsId = null;
+          pumpTtsQueue();
+          if (ttsQueue.length === 0) {
+            state = 'listening';
+            setStatus('듣는 중…');
+          }
+        }
         break;
     }
   }
@@ -269,6 +281,10 @@
       // remaining table rows: "| a | b |" -> "a, b" so cells read as a list
       .replace(/^\s*\|(.+)\|\s*$/gm, (_m, row) => row.split('|').map((c) => c.trim()).filter(Boolean).join(', '))
       .replace(/https?:\/\/\S+/g, ' ')
+      // OmniVoice chokes on a bare "(" — especially Hangul-adjacent, e.g.
+      // "영웅(Heroic)" — and stops generating audio entirely instead of just
+      // mispronouncing it. Drop the parens but keep their contents.
+      .replace(/[()（）]/g, ' ')
       .replace(/^\s{0,3}#{1,6}\s+/gm, '')
       .replace(/^\s*[-*+]\s+/gm, '')
       .replace(/^\s*>\s?/gm, '')