.

core/session.py

@@ -245,6 +245,54 @@ def _preview_text(text: str, *, limit: int = 160) -> str:
     return f"{normalized[:limit]}..."
 
 
+_KAZAKH_SPECIFIC_LETTERS = set("әғқңөұүһі")
+_RUSSIAN_SPECIFIC_LETTERS = set("ыэъё")
+
+
+def _detect_session_language(text: str) -> str | None:
+    normalized = str(text or "").strip().lower()
+    if not normalized:
+        return None
+    has_kazakh = any(ch in _KAZAKH_SPECIFIC_LETTERS for ch in normalized)
+    if has_kazakh:
+        return "kk"
+    has_russian = any(ch in _RUSSIAN_SPECIFIC_LETTERS for ch in normalized)
+    if has_russian:
+        return "ru"
+    cyrillic_chars = sum(1 for ch in normalized if "а" <= ch <= "я" or ch == "ё")
+    if cyrillic_chars >= 2:
+        return "ru"
+    return None
+
+
+_TTS_LANGUAGE_CODE_MAP = {
+    "ru": "ru",
+    "kk": "kk",
+}
+
+_STT_YANDEX_LANGUAGE_MAP = {
+    "ru": "ru-RU",
+    "kk": "kk-KZ",
+}
+
+_STT_ELEVENLABS_LANGUAGE_MAP = {
+    "ru": "rus",
+    "kk": "kaz",
+}
+
+
+def _tts_language_for(session_language: str | None) -> str | None:
+    if not session_language:
+        return None
+    return _TTS_LANGUAGE_CODE_MAP.get(session_language.lower())
+
+
+def _stt_language_for(session_language: str | None) -> str | None:
+    if not session_language:
+        return None
+    return session_language.lower() if session_language.lower() in {"ru", "kk"} else None
+
+
 def _audio_duration_ms(audio_bytes: bytes, *, sample_rate_hz: int) -> int:
     if not audio_bytes or sample_rate_hz <= 0:
         return 0
@@ -993,6 +1041,7 @@ class CallSession:
         self._personalized_greeting_template = DEFAULT_PERSONALIZED_GREETING_TEMPLATE
         self._live_stt_stream: BaseSTTStream | None = None
         self._latest_partial_transcript = ""
+        self._session_language: str | None = None
         self._default_vad_silence_timeout_ms = getattr(self._vad, "default_speech_end_silence_ms", 550)
         self._semantic_hold_silence_timeout_ms = max(self._default_vad_silence_timeout_ms, SEMANTIC_ENDPOINTING_HOLD_MS)
         LOGGER.info(
@@ -1235,6 +1284,18 @@ class CallSession:
                 )
             self._active_user_transcript = transcript
 
+            if self._session_language is None:
+                detected_language = _detect_session_language(transcript)
+                if detected_language:
+                    self._session_language = detected_language
+                    LOGGER.info(
+                        "realtime session %s language locked: epoch=%s language=%s transcript=%r",
+                        self.session_id,
+                        epoch,
+                        detected_language,
+                        _preview_text(transcript),
+                    )
+
             LOGGER.info(
                 "realtime session %s transcript accepted: epoch=%s chars=%s text=%r",
                 self.session_id,
@@ -1334,7 +1395,11 @@ class CallSession:
                 ),
                 name=f"{self.session_id}-filler-delay-{epoch}",
             )
-            async for event in self._llm.generate_stream(transcript, self._build_llm_context()):
+            async for event in self._llm.generate_stream(
+                transcript,
+                self._build_llm_context(),
+                language_code=self._session_language,
+            ):
                 self._ensure_generation(epoch)
                 if event.type == "tool_call_start":
                     LOGGER.info(
@@ -1546,7 +1611,11 @@ class CallSession:
                     await self._stt.transcribe(
                         audio_bytes,
                         keyterms=self._name_capture_keyterms() if self._awaiting_customer_name else None,
-                        language_code=self._name_capture_language_code() if self._awaiting_customer_name else None,
+                        language_code=(
+                            self._name_capture_language_code()
+                            if self._awaiting_customer_name
+                            else _stt_language_for(self._session_language)
+                        ),
                         force_batch=self._awaiting_customer_name,
                     )
                 ).strip()
@@ -1825,7 +1894,8 @@ class CallSession:
                     sentence_queue=sentence_queue,
                     started_holder=tts_started_monotonic,
                     actually_spoken_chunks=actually_spoken_chunks,
-                )
+                ),
+                language_code=_tts_language_for(self._session_language),
             ):
                 self._ensure_generation(epoch)
                 if not audio_chunk:
@@ -2061,6 +2131,7 @@ class CallSession:
             live_stt_stream = await self._stt.start_stream(
                 partial_callback=self._handle_partial_transcript,
                 keyterms=None,
+                language_code=_stt_language_for(self._session_language),
             )
         except Exception:
             LOGGER.exception("realtime session %s failed to start live STT stream", self.session_id)
@@ -2137,7 +2208,11 @@ class CallSession:
         return await self._stt.transcribe(
             utterance_audio,
             keyterms=self._name_capture_keyterms() if self._awaiting_customer_name else None,
-            language_code=self._name_capture_language_code() if self._awaiting_customer_name else None,
+            language_code=(
+                self._name_capture_language_code()
+                if self._awaiting_customer_name
+                else _stt_language_for(self._session_language)
+            ),
             force_batch=self._awaiting_customer_name,
         )