fix: implement smart filtering with dictionary hints

app.py

@@ -234,10 +234,15 @@ async def health_check():
     }
 
 # --- TRANSCRIPTION GUIDANCE ---
-# We keep prompts EMPTY to avoid "mimicry" hallucinations.
-# Language selection from the frontend is safer for large-v3.
+# We use targeted "Dictionary Hints" (keywords) to guide Whisper's spelling
+# without triggering conversational hallucinations.
 LANGUAGE_PROMPTS = {
-    "te": "", "hi": "", "ta": "", "kn": "", "ml": "", "en": ""
+    "te": "వస్తున్నాయి, విదేశీ, మెరుపులు, ఆరంభంలో, విధ్వంసం, ధన్యవాదాలు, నమస్కారం, ప్రశ్నలు, చర్చ, భారత్.",
+    "hi": "नमस्ते, क्या, क्यों, कैसे, धन्यवाद, भारत, समाचार, चर्चा।",
+    "ta": "வணக்கம், நன்றி.",
+    "kn": "ನಮಸ್ಕಾರ, ಧನ್ಯವಾದಗಳು.",
+    "ml": "നമസ്കാരം, നന്ദി.",
+    "en": ""
 }
 
 # HALLUCINATION BLACKLIST: Filter out common "filler" phrases generated by the model during silence