diff --git a/homelab-ai-bot/telegram_bot.py b/homelab-ai-bot/telegram_bot.py
index 536c5bf5..9cd1694e 100644
--- a/homelab-ai-bot/telegram_bot.py
+++ b/homelab-ai-bot/telegram_bot.py
@@ -460,6 +460,10 @@ async def handle_voice(update: Update, ctx: ContextTypes.DEFAULT_TYPE):
         document_mode = doc_prefix or rag_mode.is_document_mode(channel_key)
 
         log.info("Voice transkribiert: %s", text[:100])
+        prev = ACTIVE_LLM_TASKS.pop(update.effective_chat.id, None)
+        if prev and not prev.done():
+            prev.cancel()
+            log.info("Vorheriger LLM-Lauf abgebrochen (neue Sprachnachricht)")
         await update.message.reply_text(f"🗣 \"{text}\"\n\n🤔 Denke nach...")
 
         session_id = memory_client.get_or_create_session(channel_key, source="telegram")
@@ -479,11 +483,14 @@ async def handle_voice(update: Update, ctx: ContextTypes.DEFAULT_TYPE):
         ACTIVE_LLM_TASKS[update.effective_chat.id] = llm_task
 
         waited = 0
+        # Kurze Intervalle: bei Save.TV/LLM wirkt 30s ohne Nachricht wie „Haenger“
         while not llm_task.done():
-            await asyncio.sleep(30)
-            waited += 30
+            await asyncio.sleep(10)
+            waited += 10
             if not llm_task.done():
-                await update.message.reply_text("⏳ Suche laeuft noch (" + str(waited) + "s)...")
+                await update.message.reply_text(
+                    "⏳ Noch dran (" + str(waited) + "s) — Save.TV/Modell kann etwas brauchen…"
+                )
 
         answer = await llm_task
 
@@ -845,6 +852,10 @@ async def handle_message(update: Update, ctx: ContextTypes.DEFAULT_TYPE):
         return
     document_mode = doc_prefix or rag_mode.is_document_mode(channel_key)
 
+    prev = ACTIVE_LLM_TASKS.pop(update.effective_chat.id, None)
+    if prev and not prev.done():
+        prev.cancel()
+        log.info("Vorheriger LLM-Lauf abgebrochen (neue Freitext-Nachricht)")
     await update.message.reply_text("🤔 Denke nach...")
     if _likely_deep_research_request(work_text):
         await update.message.reply_text("🔎 Deep Research gestartet. Das dauert meist 2-5 Minuten.")
@@ -867,10 +878,12 @@ async def handle_message(update: Update, ctx: ContextTypes.DEFAULT_TYPE):
 
         waited = 0
         while not llm_task.done():
-            await asyncio.sleep(30)
-            waited += 30
+            await asyncio.sleep(10)
+            waited += 10
             if not llm_task.done():
-                await update.message.reply_text("Suche laeuft noch (" + str(waited) + "s)...")
+                await update.message.reply_text(
+                    "⏳ Noch dran (" + str(waited) + "s) — Save.TV/Modell kann etwas brauchen…"
+                )
 
         answer = await llm_task