.

callback_client.py

@@ -0,0 +1,120 @@
+from __future__ import annotations
+
+import logging
+import os
+from typing import Any
+
+import httpx
+
+from realtime_voice_service.crm_client import _issue_service_token
+
+
+LOGGER = logging.getLogger("uvicorn.error")
+
+
+def _enabled() -> bool:
+    return os.getenv("CALLBACK_DISPATCH_ENABLED", "1").strip().lower() in {"1", "true", "yes"}
+
+
+def _bridge_base_url() -> str:
+    return str(
+        os.getenv("ASTERISK_BRIDGE_SERVICE_URL", "http://asterisk-bridge-service:8000")
+    ).rstrip("/")
+
+
+def _secret() -> str:
+    return str(os.getenv("CRM_APP_TOKEN_SECRET", "")).strip()
+
+
+def _request_headers() -> dict[str, str] | None:
+    secret = _secret()
+    if not secret:
+        return None
+    return {"Authorization": f"Bearer {_issue_service_token(secret)}"}
+
+
+async def schedule_callback(
+    *,
+    phone: str,
+    in_minutes: int | None = None,
+    scheduled_for: str | None = None,
+    reason: str | None = None,
+    customer_name: str | None = None,
+    source_call_id: str | None = None,
+    source_interaction_id: str | None = None,
+    idempotency_key: str | None = None,
+) -> dict[str, Any] | None:
+    if not _enabled():
+        LOGGER.info("callback dispatch disabled by env")
+        return None
+
+    headers = _request_headers()
+    if headers is None:
+        LOGGER.warning("callback client: CRM_APP_TOKEN_SECRET is not configured")
+        return None
+
+    payload: dict[str, Any] = {"phone": phone}
+    if in_minutes is not None:
+        payload["in_minutes"] = int(in_minutes)
+    if scheduled_for:
+        payload["scheduled_for"] = scheduled_for
+    if reason:
+        payload["reason"] = reason
+    if customer_name:
+        payload["customer_name"] = customer_name
+    if source_call_id:
+        payload["source_call_id"] = source_call_id
+    if source_interaction_id:
+        payload["source_interaction_id"] = source_interaction_id
+    if idempotency_key:
+        payload["idempotency_key"] = idempotency_key
+
+    url = f"{_bridge_base_url()}/asterisk/callbacks"
+    try:
+        async with httpx.AsyncClient(timeout=8.0) as client:
+            resp = await client.post(url, json=payload, headers=headers)
+            if resp.status_code in (200, 201):
+                data = resp.json()
+                LOGGER.info(
+                    "callback scheduled: job_id=%s phone=%s scheduled_for=%s",
+                    data.get("job_id"),
+                    data.get("phone_e164"),
+                    data.get("scheduled_for"),
+                )
+                return data
+            if resp.status_code == 409:
+                LOGGER.info("callback already scheduled for phone=%s: %s", phone, resp.text[:200])
+                return {"status": "conflict", "detail": resp.text}
+            LOGGER.warning(
+                "callback schedule failed: status=%s body=%s",
+                resp.status_code,
+                resp.text[:300],
+            )
+            return None
+    except Exception:
+        LOGGER.exception("callback client: request failed phone=%s", phone)
+        return None
+
+
+async def cancel_callback(*, job_id: str) -> dict[str, Any] | None:
+    if not _enabled():
+        return None
+    headers = _request_headers()
+    if headers is None:
+        return None
+    url = f"{_bridge_base_url()}/asterisk/callbacks/{job_id}"
+    try:
+        async with httpx.AsyncClient(timeout=8.0) as client:
+            resp = await client.delete(url, headers=headers)
+            if resp.status_code == 200:
+                return resp.json()
+            LOGGER.warning(
+                "callback cancel failed: job_id=%s status=%s body=%s",
+                job_id,
+                resp.status_code,
+                resp.text[:300],
+            )
+            return None
+    except Exception:
+        LOGGER.exception("callback client: cancel failed job_id=%s", job_id)
+        return None

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,
         )
 

providers/base.py

@@ -47,20 +47,32 @@ class BaseSTT(ABC):
         *,
         partial_callback: PartialTranscriptCallback | None = None,
         keyterms: Sequence[str] | None = None,
+        language_code: str | None = None,
     ) -> BaseSTTStream | None:
-        del partial_callback, keyterms
+        del partial_callback, keyterms, language_code
         return None
 
 
 class BaseLLM(ABC):
     @abstractmethod
-    async def generate_stream(self, text: str, context: list) -> AsyncGenerator[LLMStreamEvent, None]:
+    async def generate_stream(
+        self,
+        text: str,
+        context: list,
+        *,
+        language_code: str | None = None,
+    ) -> AsyncGenerator[LLMStreamEvent, None]:
         raise NotImplementedError
 
 
 class BaseTTS(ABC):
     @abstractmethod
-    async def synthesize_stream(self, text_stream: AsyncIterable[str]) -> AsyncGenerator[bytes, None]:
+    async def synthesize_stream(
+        self,
+        text_stream: AsyncIterable[str],
+        *,
+        language_code: str | None = None,
+    ) -> AsyncGenerator[bytes, None]:
         raise NotImplementedError
 
 
@@ -106,8 +118,9 @@ class MockSTT(BaseSTT):
         *,
         partial_callback: PartialTranscriptCallback | None = None,
         keyterms: Sequence[str] | None = None,
+        language_code: str | None = None,
     ) -> BaseSTTStream | None:
-        del keyterms
+        del keyterms, language_code
         return _MockSTTStream(parent=self, partial_callback=partial_callback)
 
 
@@ -121,8 +134,14 @@ class MockLLM(BaseLLM):
         self._token_delay_ms = max(token_delay_ms, 0)
         self._scripted_responses = list(scripted_responses or [])
 
-    async def generate_stream(self, text: str, context: list) -> AsyncGenerator[LLMStreamEvent, None]:
+    async def generate_stream(
-        del context
+        self,
+        text: str,
+        context: list,
+        *,
+        language_code: str | None = None,
+    ) -> AsyncGenerator[LLMStreamEvent, None]:
+        del context, language_code
         response_text = (
             self._scripted_responses.pop(0)
             if self._scripted_responses
@@ -152,7 +171,13 @@ class MockTTS(BaseTTS):
         self._amplitude = amplitude
         self._milliseconds_per_word = max(milliseconds_per_word, 40)
 
-    async def synthesize_stream(self, text_stream: AsyncIterable[str]) -> AsyncGenerator[bytes, None]:
+    async def synthesize_stream(
+        self,
+        text_stream: AsyncIterable[str],
+        *,
+        language_code: str | None = None,
+    ) -> AsyncGenerator[bytes, None]:
+        del language_code
         async for text in text_stream:
             normalized = text.strip()
             if not normalized:

providers/llm.py

@@ -31,6 +31,45 @@ _WEB_SEARCH_POLICY = (
     "Если для поиска не хватает города, даты или объекта, задай короткий уточняющий вопрос."
 )
 _WEB_SEARCH_POLICY_MARKER = "если клиент спрашивает про погоду"
+_CALLBACK_POLICY = (
+    "Правило перезвона: если клиент просит перезвонить позже (через N минут, в определённое время или просит "
+    "связаться с ним позже), вызови инструмент schedule_callback с номером телефона клиента и временем. "
+    "Если номер не назван — попроси клиента продиктовать его. После успешного вызова инструмента подтверди "
+    "клиенту коротко на русском, в какое время мы перезвоним. Не вызывай инструмент несколько раз подряд "
+    "и не обещай перезвон до того, как инструмент вернул успех."
+)
+_CALLBACK_POLICY_MARKER = "если клиент просит перезвонить позже"
+_LANGUAGE_DIRECTIVES = {
+    "ru": (
+        "Язык диалога: русский. Отвечай только на русском языке, естественно и кратко, в одном спокойном "
+        "стиле для голоса. Не переключайся на другие языки, даже если в запросе клиента есть отдельные "
+        "слова на другом языке."
+    ),
+    "kk": (
+        "Сөйлесу тілі: қазақ тілі. Тек қазақ тілінде, табиғи және қысқа жауап бер, дауыс үшін бір қалыпты "
+        "стильде. Клиенттің сөзінде бөтен тілден бөлек сөздер болса да, басқа тілге ауыспа."
+    ),
+}
+
+
+def _normalize_language_code(value: str | None) -> str | None:
+    if value is None:
+        return None
+    raw = str(value).strip().lower().replace("_", "-")
+    if not raw:
+        return None
+    if raw in {"ru", "rus", "ru-ru"}:
+        return "ru"
+    if raw in {"kk", "kaz", "kz", "kk-kz", "kz-kz"}:
+        return "kk"
+    return raw[:2] or None
+
+
+def _language_directive(language_code: str | None) -> str | None:
+    normalized = _normalize_language_code(language_code)
+    if normalized is None:
+        return None
+    return _LANGUAGE_DIRECTIVES.get(normalized)
 
 _LIVE_SEARCH_PATTERN = re.compile(
     r"("
@@ -72,9 +111,11 @@ def _with_conversation_close_policy(system_prompt: str) -> str:
 
 def _with_runtime_policies(system_prompt: str) -> str:
     prompt = _with_conversation_close_policy(system_prompt)
-    if _WEB_SEARCH_POLICY_MARKER in prompt:
+    if _WEB_SEARCH_POLICY_MARKER not in prompt:
-        return prompt
+        prompt = f"{prompt}\n\n{_WEB_SEARCH_POLICY}" if prompt else _WEB_SEARCH_POLICY
-    return f"{prompt}\n\n{_WEB_SEARCH_POLICY}" if prompt else _WEB_SEARCH_POLICY
+    if _CALLBACK_POLICY_MARKER not in prompt:
+        prompt = f"{prompt}\n\n{_CALLBACK_POLICY}" if prompt else _CALLBACK_POLICY
+    return prompt
 
 
 def _should_force_live_search(text: str) -> bool:
@@ -198,11 +239,17 @@ class OpenAILLM(BaseLLM):
             self._max_tool_roundtrips,
         )
 
-    async def generate_stream(self, text: str, context: list) -> AsyncGenerator[LLMStreamEvent, None]:
+    async def generate_stream(
+        self,
+        text: str,
+        context: list,
+        *,
+        language_code: str | None = None,
+    ) -> AsyncGenerator[LLMStreamEvent, None]:
         if not self._api_key:
             raise RuntimeError("OPENAI_API_KEY is required for OpenAI LLM")
 
-        messages = self._build_messages(text, context)
+        messages = self._build_messages(text, context, language_code=language_code)
         force_live_search = _should_force_live_search(text)
         turn_started_monotonic = time.perf_counter()
         LOGGER.info(
@@ -478,6 +525,8 @@ class OpenAILLM(BaseLLM):
         )
         if name == "serper":
             result = await self._run_serper_tool(arguments)
+        elif name == "schedule_callback":
+            result = await self._run_schedule_callback_tool(arguments, tool_call_id=str(tool_call.get("id") or ""))
         else:
             result = f"Tool `{name}` is not supported by this runtime."
         LOGGER.info(
@@ -547,6 +596,64 @@ class OpenAILLM(BaseLLM):
         LOGGER.info("Serper summary built: chars=%s preview=%r", len(summary), _preview_text(summary))
         return summary
 
+    async def _run_schedule_callback_tool(self, arguments: dict[str, Any], *, tool_call_id: str) -> str:
+        from realtime_voice_service import callback_client
+
+        phone = str(arguments.get("phone") or "").strip()
+        if not phone:
+            return "Не указан номер телефона. Попроси клиента продиктовать номер."
+
+        in_minutes = arguments.get("in_minutes")
+        scheduled_for = arguments.get("scheduled_for")
+        if in_minutes is None and not scheduled_for:
+            return "Не указано время: передай in_minutes или scheduled_for."
+
+        try:
+            in_minutes_int = int(in_minutes) if in_minutes is not None else None
+        except (TypeError, ValueError):
+            in_minutes_int = None
+
+        reason = str(arguments.get("reason") or "").strip() or None
+        customer_name = str(arguments.get("customer_name") or "").strip() or None
+        idempotency_key = (tool_call_id or "")[:128] or None
+
+        LOGGER.info(
+            "schedule_callback tool start: phone=%s in_minutes=%s scheduled_for=%s reason=%r",
+            phone,
+            in_minutes_int,
+            scheduled_for,
+            _preview_text(reason or ""),
+        )
+
+        result = await callback_client.schedule_callback(
+            phone=phone,
+            in_minutes=in_minutes_int,
+            scheduled_for=str(scheduled_for) if scheduled_for else None,
+            reason=reason,
+            customer_name=customer_name,
+            idempotency_key=idempotency_key,
+        )
+
+        if result is None:
+            return (
+                "Не удалось запланировать перезвон: сервис недоступен. Извинись и предложи "
+                "клиенту перезвонить самостоятельно."
+            )
+        if result.get("status") == "conflict":
+            return (
+                "На этот номер уже запланирован активный перезвон. Сообщи клиенту, что "
+                "перезвон уже в очереди."
+            )
+
+        scheduled = result.get("scheduled_for") or scheduled_for or "soon"
+        phone_e164 = result.get("phone_e164") or phone
+        job_id = result.get("job_id") or ""
+        return (
+            f"Перезвон запланирован: job_id={job_id}, номер {phone_e164}, время {scheduled} (UTC). "
+            "Подтверди клиенту коротко на русском, в какое примерно время мы перезвоним "
+            "(переведи UTC во время клиента, если знаешь его часовой пояс — иначе скажи в минутах от сейчас)."
+        )
+
     async def _get_serper_session(self):
         try:
             import aiohttp
@@ -564,46 +671,96 @@ class OpenAILLM(BaseLLM):
             return self._serper_session
 
     def _build_tools(self) -> list[dict[str, Any]] | None:
-        if not self._enable_tools or not self._serper_api_key:
+        if not self._enable_tools:
             return None
-        return [
+        tools: list[dict[str, Any]] = []
-            {
+        if self._serper_api_key:
-                "type": "function",
+            tools.append(
-                "function": {
+                {
-                    "name": "serper",
+                    "type": "function",
-                    "description": (
+                    "function": {
-                        "Search the public web for recent or external information when the user asks "
+                        "name": "serper",
-                        "about weather, current facts, news, prices, currency rates, websites, company data, "
+                        "description": (
-                        "schedules, addresses, phone numbers, or anything requiring live search."
+                            "Search the public web for recent or external information when the user asks "
-                    ),
+                            "about weather, current facts, news, prices, currency rates, websites, company data, "
-                    "parameters": {
+                            "schedules, addresses, phone numbers, or anything requiring live search."
-                        "type": "object",
+                        ),
-                        "properties": {
+                        "parameters": {
-                            "query": {
+                            "type": "object",
-                                "type": "string",
+                            "properties": {
-                                "description": "Precise search query to send to Serper.",
+                                "query": {
-                            },
+                                    "type": "string",
-                            "num": {
+                                    "description": "Precise search query to send to Serper.",
-                                "type": "integer",
+                                },
-                                "description": "How many results to fetch, usually 3 to 5.",
+                                "num": {
-                                "minimum": 1,
+                                    "type": "integer",
-                                "maximum": 10,
+                                    "description": "How many results to fetch, usually 3 to 5.",
-                            },
+                                    "minimum": 1,
-                            "hl": {
+                                    "maximum": 10,
-                                "type": "string",
+                                },
-                                "description": "UI language code, for example ru or en.",
+                                "hl": {
-                            },
+                                    "type": "string",
-                            "gl": {
+                                    "description": "UI language code, for example ru or en.",
-                                "type": "string",
+                                },
-                                "description": "Country code for result localization, for example kz or us.",
+                                "gl": {
+                                    "type": "string",
+                                    "description": "Country code for result localization, for example kz or us.",
+                                },
                             },
+                            "required": ["query"],
+                            "additionalProperties": False,
                         },
-                        "required": ["query"],
-                        "additionalProperties": False,
                     },
-                },
+                }
-            }
+            )
-        ]
+        if _read_bool_env("CALLBACK_DISPATCH_ENABLED", True):
+            tools.append(
+                {
+                    "type": "function",
+                    "function": {
+                        "name": "schedule_callback",
+                        "description": (
+                            "Schedule an outbound callback to the customer when they ask to be called back later. "
+                            "The system will dial the customer at the requested time and reconnect them with the AI. "
+                            "Use only when the customer explicitly requests a callback. After the tool succeeds, "
+                            "confirm the callback time briefly in Russian."
+                        ),
+                        "parameters": {
+                            "type": "object",
+                            "properties": {
+                                "phone": {
+                                    "type": "string",
+                                    "description": (
+                                        "Customer phone number. Accepts +7XXXXXXXXXX, 8XXXXXXXXXX, or any "
+                                        "format with digits — the backend normalizes it."
+                                    ),
+                                },
+                                "in_minutes": {
+                                    "type": "integer",
+                                    "description": "How many minutes from now to call back. Use either in_minutes or scheduled_for.",
+                                    "minimum": 0,
+                                    "maximum": 10080,
+                                },
+                                "scheduled_for": {
+                                    "type": "string",
+                                    "description": "ISO-8601 UTC datetime to dial at, e.g. 2026-05-02T14:30:00Z. Use either this or in_minutes.",
+                                },
+                                "reason": {
+                                    "type": "string",
+                                    "description": "One short sentence in Russian: why the callback is needed (the customer's topic).",
+                                },
+                                "customer_name": {
+                                    "type": "string",
+                                    "description": "Customer's name if already known from the conversation.",
+                                },
+                            },
+                            "required": ["phone"],
+                            "additionalProperties": False,
+                        },
+                    },
+                }
+            )
+        return tools or None
 
     def _get_client(self):
         if self._client is not None:
@@ -633,10 +790,19 @@ class OpenAILLM(BaseLLM):
     def _supports_custom_temperature(self) -> bool:
         return not self._model.lower().startswith("gpt-5")
 
-    def _build_messages(self, text: str, context: list) -> list[dict[str, Any]]:
+    def _build_messages(
+        self,
+        text: str,
+        context: list,
+        *,
+        language_code: str | None = None,
+    ) -> list[dict[str, Any]]:
         messages: list[dict[str, Any]] = []
         if self._system_prompt:
             messages.append({"role": "system", "content": self._system_prompt})
+        language_directive = _language_directive(language_code)
+        if language_directive:
+            messages.append({"role": "system", "content": language_directive})
 
         context_messages: list[dict[str, Any]] = []
         for entry in context:
@@ -780,8 +946,14 @@ class OllamaLLM(BaseLLM):
             self._num_ctx or "default",
         )
 
-    async def generate_stream(self, text: str, context: list) -> AsyncGenerator[LLMStreamEvent, None]:
+    async def generate_stream(
-        messages = self._build_messages(text, context)
+        self,
+        text: str,
+        context: list,
+        *,
+        language_code: str | None = None,
+    ) -> AsyncGenerator[LLMStreamEvent, None]:
+        messages = self._build_messages(text, context, language_code=language_code)
         options: dict[str, Any] = {"temperature": self._temperature}
         if self._num_predict > 0:
             options["num_predict"] = self._num_predict
@@ -874,10 +1046,19 @@ class OllamaLLM(BaseLLM):
         self._client = httpx.AsyncClient(timeout=timeout)
         return self._client
 
-    def _build_messages(self, text: str, context: list) -> list[dict[str, Any]]:
+    def _build_messages(
+        self,
+        text: str,
+        context: list,
+        *,
+        language_code: str | None = None,
+    ) -> list[dict[str, Any]]:
         messages: list[dict[str, Any]] = []
         if self._system_prompt:
             messages.append({"role": "system", "content": self._system_prompt})
+        language_directive = _language_directive(language_code)
+        if language_directive:
+            messages.append({"role": "system", "content": language_directive})
 
         context_messages: list[dict[str, Any]] = []
         for entry in context:

providers/stt.py

@@ -663,15 +663,18 @@ class ElevenLabsSTT(BaseSTT):
         *,
         partial_callback: PartialTranscriptCallback | None = None,
         keyterms: Sequence[str] | None = None,
+        language_code: str | None = None,
     ) -> BaseSTTStream | None:
         if not self._use_realtime:
             return None
         if not self._api_key:
             raise RuntimeError("ELEVENLABS_API_KEY is required for ElevenLabs STT")
 
+        effective_language_code = _normalize_elevenlabs_language_code(language_code) or self._language_code
         websocket_url = self._build_realtime_websocket_url(
             sample_rate_hz=self._target_sample_rate_hz,
             keyterms=keyterms,
+            language_code=effective_language_code,
         )
         normalized_keyterm_count = len(
             _normalize_keyterms(
@@ -684,7 +687,7 @@ class ElevenLabsSTT(BaseSTT):
             "ElevenLabs STT live stream starting: realtime_model=%s sample_rate=%s language=%s keyterms=%s",
             self._realtime_model_id,
             self._target_sample_rate_hz,
-            self._language_code,
+            effective_language_code,
             normalized_keyterm_count,
         )
         stream = ElevenLabsRealtimeSTTStream(
@@ -1211,6 +1214,33 @@ class FallbackSTT(BaseSTT):
                 force_batch=force_batch,
             )
 
+    async def start_stream(
+        self,
+        *,
+        partial_callback: PartialTranscriptCallback | None = None,
+        keyterms: Sequence[str] | None = None,
+        language_code: str | None = None,
+    ) -> BaseSTTStream | None:
+        try:
+            stream = await self._primary.start_stream(
+                partial_callback=partial_callback,
+                keyterms=keyterms,
+                language_code=language_code,
+            )
+            if stream is not None:
+                return stream
+        except Exception:
+            LOGGER.exception(
+                "Primary STT start_stream failed; falling back: primary=%s fallback=%s",
+                type(self._primary).__name__,
+                type(self._fallback).__name__,
+            )
+        return await self._fallback.start_stream(
+            partial_callback=partial_callback,
+            keyterms=keyterms,
+            language_code=language_code,
+        )
+
     async def close(self) -> None:
         for provider in (self._primary, self._fallback):
             close = getattr(provider, "close", None)

providers/tts.py

@@ -166,7 +166,12 @@ class ElevenLabsTTS(BaseTTS):
             self._voice_settings,
         )
 
-    async def synthesize_stream(self, text_stream: AsyncIterable[str]) -> AsyncGenerator[bytes, None]:
+    async def synthesize_stream(
+        self,
+        text_stream: AsyncIterable[str],
+        *,
+        language_code: str | None = None,
+    ) -> AsyncGenerator[bytes, None]:
         if not self._api_key:
             raise RuntimeError("ELEVENLABS_API_KEY is required for ElevenLabs TTS")
         if not self._voice_id:
@@ -178,7 +183,8 @@ class ElevenLabsTTS(BaseTTS):
         except Exception as exc:  # noqa: BLE001
             raise RuntimeError("The `websockets` package is required for ElevenLabs TTS WebSocket streaming") from exc
 
-        websocket_url = self._build_websocket_url()
+        effective_language_code = (str(language_code).strip() if language_code else "") or self._language_code
+        websocket_url = self._build_websocket_url(language_code=effective_language_code)
         started_monotonic = time.perf_counter()
         audio_chunk_count = 0
         audio_byte_count = 0
@@ -193,7 +199,7 @@ class ElevenLabsTTS(BaseTTS):
             self._output_format,
             self._provider_sample_rate_hz,
             self._target_sample_rate_hz,
-            self._language_code,
+            effective_language_code,
             self._auto_mode,
         )
         try:
@@ -290,7 +296,7 @@ class ElevenLabsTTS(BaseTTS):
         except OSError as exc:
             raise RuntimeError("ElevenLabs TTS WebSocket connection failed") from exc
 
-    def _build_websocket_url(self) -> str:
+    def _build_websocket_url(self, *, language_code: str | None = None) -> str:
         query = {
             "model_id": self._model_id,
             "output_format": self._output_format,
@@ -299,8 +305,9 @@ class ElevenLabsTTS(BaseTTS):
             "sync_alignment": "false",
             "apply_text_normalization": "auto",
         }
-        if self._language_code:
+        effective_language = (str(language_code).strip() if language_code else "") or self._language_code
-            query["language_code"] = self._language_code
+        if effective_language:
+            query["language_code"] = effective_language
         encoded_voice_id = quote(self._voice_id, safe="")
         return f"{self._ws_base}/v1/text-to-speech/{encoded_voice_id}/stream-input?{urlencode(query)}"