corrected language detection

python-services/llm_server.py

@@ -20,7 +20,8 @@ from pydantic import BaseModel
 # Internal Domain Layer Imports
 from client import groq, fireworks
 from conversation_manager import ConversationManager, ConversationSession
-from language_config import normalize_language, SUPPORTED_LANGUAGES
+from utils import detect_language_from_content, SUPPORTED_LANGUAGES
+from language_config import normalize_language
 from prompt_builder import build_system_prompt
 from summary import generate_summary
 from caller_info_extractor import extract_name_from_text, extract_phone_from_text
@@ -122,37 +123,47 @@ def _extract_sentences(buffer: str) -> tuple[list[str], str]:
     return sentences, buffer[last_end:]
 
 
-async def _fire_tts(sentence: str, language: str, gender: str) -> Optional[bytes]:
+async def _fire_tts(sentence: str, language: str, gender: str) -> tuple[Optional[bytes], str]:
     """
     Asynchronously POST a single sentence to the deployed TTS Space.
-    Endpoint: POST {TTS_SPACE_URL}/synthesise
-    Payload:  {"text": str, "language": str, "gender": str}
-    Returns raw audio bytes (audio/mpeg) on success, None on any failure.
+    Language is resolved from sentence content so Urdu/English never collide.
+    Returns (audio_bytes, tts_language) on success, (None, tts_language) on failure.
     """
     sentence = sentence.strip()
     if not sentence:
-        return None
+        return None, normalize_language(language)
+
+    tts_lang = normalize_language(detect_language_from_content(sentence))
+    session_lang = normalize_language(language)
+    if tts_lang != session_lang:
+        logger.info(
+            "TTS lang from content: session=%s → sentence=%s preview=%r",
+            session_lang,
+            tts_lang,
+            sentence[:60],
+        )
+
     try:
         async with httpx.AsyncClient(timeout=TTS_SYNTHESISE_TIMEOUT) as client:
             resp = await client.post(
                 f"{TTS_SPACE_URL}/synthesise",
-                json={"text": sentence, "language": language, "gender": gender},
+                json={"text": sentence, "language": tts_lang, "gender": gender},
             )
         if resp.status_code == 200:
-            logger.info(f"TTS ✓  lang={language}  chars={len(sentence)}")
-            return resp.content
+            logger.info(f"TTS ✓  lang={tts_lang}  chars={len(sentence)}")
+            return resp.content, tts_lang
         logger.warning(
             f"TTS Space returned HTTP {resp.status_code}: {resp.text[:120]}"
         )
-        return None
+        return None, tts_lang
     except httpx.TimeoutException:
         logger.error(
             f"TTS request timed out ({TTS_SYNTHESISE_TIMEOUT}s): {sentence[:60]!r}"
         )
-        return None
+        return None, tts_lang
     except Exception as exc:
         logger.error(f"TTS request error: {exc}")
-        return None
+        return None, tts_lang
 
 
 def _start_groq_stream_thread(
@@ -534,13 +545,13 @@ async def chat_stream(req: ChatRequest):
             Fire TTS for one sentence.
             Returns a ready-to-yield SSE audio event string, or None if TTS failed.
             """
-            audio_bytes = await _fire_tts(sentence, detected_language, tts_gender)
+            audio_bytes, tts_lang = await _fire_tts(sentence, detected_language, tts_gender)
             if audio_bytes:
                 return _sse({
                     "type": "audio",
                     "sentence": sentence,
                     "audio_b64": base64.b64encode(audio_bytes).decode("utf-8"),
-                    "lang": detected_language,
+                    "lang": tts_lang,
                 })
             return None
 

python-services/utils.py

@@ -12,6 +12,8 @@ from language_config import (
     GURMUKHI,
     has_arabic_script,
     normalize_language,
+    SUPPORTED_LANGUAGES,
+    URDU_SHARED_PHRASES,
 )
 
 
@@ -45,6 +47,11 @@ def detect_language_from_content(text: str) -> str:
     if detect_roman_urdu(text):
         return "ur"
 
+    text_lower = text.lower()
+    for phrase in URDU_SHARED_PHRASES:
+        if phrase.lower() in text_lower:
+            return "ur"
+
     return "en"