fix: handle farewells and stabilize voice responses

.env.example

@@ -9,7 +9,7 @@ LLM_PROVIDER=openai
 OPENAI_API_KEY=
 OPENAI_BASE_URL=
 OPENAI_LLM_MODEL=gpt-4o-mini
-OPENAI_LLM_SYSTEM_PROMPT=You are Ainur, a concise Russian-speaking voice agent for the DigiOps contact center. Always answer in Russian, naturally and briefly for voice playback. If the caller asks whether you are a robot, AI, or human, answer exactly: "Да, я голосовой агент контакт-центра DigiOps." Read phone numbers digit by digit. Prefer clear stress-friendly wording. Do not use markdown, URLs, or tables.
+OPENAI_LLM_SYSTEM_PROMPT=You are Ainur, a concise Russian-speaking voice agent for the DigiOps contact center. Always answer in Russian, naturally, briefly, and in one consistent speaking style for voice playback. If the caller asks whether you are a robot, AI, or human, answer exactly: "Да, я голосовой агент контакт-центра DigiOps." If the caller says goodbye or thanks and ends the conversation, respond briefly and politely without filler phrases. Read phone numbers digit by digit. Prefer clear stress-friendly wording. Do not use markdown, URLs, or tables.
 OPENAI_TIMEOUT_SECONDS=30
 OPENAI_STT_MODEL=whisper-1
 OPENAI_STT_LANGUAGE=ru
@@ -26,7 +26,7 @@ OLLAMA_LLM_NUM_CTX=1024
 
 STT_PROVIDER=openai
 STT_FALLBACK_PROVIDER=elevenlabs
-STT_PROMPT=Это разговор на русском языке, но могут встречаться казахские имена: Айбын, Магжан, Бауыржан, Асель.
+STT_PROMPT=Это русская речь. Точно распознавай имена, фамилии, слова благодарности и прощания. Возможны имена: Айнур, Айгерим, Алия, Данияр, Азамат, Мадина, Нурсултан. Номера телефонов записывай цифрами.
 TTS_PROVIDER=elevenlabs
 
 ENABLE_AUDIO_DUMP=false
@@ -54,12 +54,12 @@ YANDEX_STT_TOPIC=general
 
 VAD_THRESHOLD=0.5
 VAD_NEGATIVE_THRESHOLD=
-VAD_SILENCE_TIMEOUT_MS=350
-VAD_SPEECH_PAD_MS=64
+VAD_SILENCE_TIMEOUT_MS=420
+VAD_SPEECH_PAD_MS=120
 VAD_MIN_SPEECH_DURATION_MS=0
 VAD_USE_ONNX=false
 REALTIME_VOICE_INITIAL_GREETING_TEXT=Здравствуйте! Меня зовут Айнур, я эй-ай-ассистент компании ДиджиОпс. Как я могу к вам обращаться?
 SEMANTIC_ENDPOINTING_HOLD_MS=600
-REALTIME_VOICE_FILLER_DELAY_MS=250
+REALTIME_VOICE_FILLER_DELAY_MS=200
 REALTIME_VOICE_TTS_CHUNK_SOFT_MIN_CHARS=10
 REALTIME_VOICE_TTS_CHUNK_SOFT_MIN_WORDS=2

core/filler_audio.py

@@ -17,11 +17,11 @@ LOGGER = logging.getLogger("uvicorn.error")
 DEFAULT_FILLER_TEXTS: dict[str, tuple[str, ...]] = {
     "serper": (
         "Сейчас уточню.",
-        "Минуту, пожалуйста.",
+        "Проверяю информацию.",
     ),
     "generic": (
-        "Секунду.",
-        "Минуту.",
+        "Сейчас отвечу.",
+        "Уточняю.",
     ),
 }
 

core/session.py

@@ -52,6 +52,49 @@ SEMANTIC_ENDPOINTING_HOLD_MS = _semantic_hold_ms()
 FILLER_AUDIO_DELAY_MS = _filler_delay_ms()
 TTS_CHUNK_SOFT_MIN_CHARS = _tts_chunk_soft_min_chars()
 TTS_CHUNK_SOFT_MIN_WORDS = _tts_chunk_soft_min_words()
+_FAREWELL_MARKERS: frozenset[str] = frozenset({
+    "спасибо",
+    "благодарю",
+    "до свидания",
+    "всего доброго",
+    "пока",
+    "до встречи",
+    "хорошего дня",
+    "goodbye",
+    "bye",
+})
+_FAREWELL_QUERY_MARKERS: frozenset[str] = frozenset({
+    "вопрос",
+    "подскаж",
+    "скажи",
+    "почему",
+    "как",
+    "что",
+    "где",
+    "когда",
+    "можно",
+    "нужно",
+    "хочу",
+    "мне",
+})
+FAREWELL_RESPONSE_TEXT = "Спасибо за обращение в DigiOps. Всего доброго, до свидания!"
+
+
+def _is_farewell_transcript(text: str) -> bool:
+    normalized = _voice_text_key(text)
+    return any(marker in normalized for marker in _FAREWELL_MARKERS)
+
+
+def _is_terminal_farewell(text: str) -> bool:
+    normalized = _voice_text_key(text)
+    if not any(marker in normalized for marker in _FAREWELL_MARKERS):
+        return False
+    if any(marker in normalized for marker in _FAREWELL_QUERY_MARKERS):
+        return False
+    word_count = len(re.findall(r"[^\W\d_]+(?:[-'][^\W\d_]+)*", normalized, flags=re.UNICODE))
+    return word_count <= 10
+
+
 SEMANTIC_CONTINUATION_TOKENS = {
     "а",
     "в",
@@ -927,6 +970,20 @@ class CallSession:
                 await self._assistant_task
         await self.transport.close()
 
+    async def _close_after_farewell(self) -> None:
+        if self._closed:
+            return
+        self._closed = True
+        LOGGER.info(
+            "realtime session %s closing transport after farewell: state=%s epoch=%s conversation_entries=%s",
+            self.session_id,
+            self.state.value,
+            self.generation_epoch,
+            len(self._conversation),
+        )
+        await self._cancel_live_stt_stream()
+        await self.transport.close()
+
     def _start_assistant_turn(
         self,
         *,
@@ -1047,6 +1104,26 @@ class CallSession:
                 )
                 self._set_state(SessionState.LISTENING, reason="name collection completed")
                 return
+            if _is_terminal_farewell(transcript):
+                LOGGER.info(
+                    "realtime session %s farewell response prepared: epoch=%s text=%r",
+                    self.session_id,
+                    epoch,
+                    _preview_text(FAREWELL_RESPONSE_TEXT),
+                )
+                await self._play_prepared_response(
+                    epoch=epoch,
+                    response_text=FAREWELL_RESPONSE_TEXT,
+                    actually_spoken_chunks=actually_spoken_chunks,
+                )
+                self._ensure_generation(epoch)
+                self._commit_assistant_context(
+                    llm_generated_text=FAREWELL_RESPONSE_TEXT,
+                    actually_spoken_chunks=actually_spoken_chunks,
+                    interrupted=False,
+                )
+                await self._close_after_farewell()
+                return
             sentence_queue = asyncio.Queue()
             self._sentence_queue = sentence_queue
             playback_task = asyncio.create_task(
@@ -1066,6 +1143,7 @@ class CallSession:
                     epoch=epoch,
                     tool_name=None,
                     started_monotonic=llm_started_monotonic,
+                    suppress=_is_farewell_transcript(transcript),
                 ),
                 name=f"{self.session_id}-filler-delay-{epoch}",
             )
@@ -1618,7 +1696,10 @@ class CallSession:
         epoch: int,
         tool_name: str | None,
         started_monotonic: float,
+        suppress: bool = False,
     ) -> None:
+        if suppress:
+            return
         if FILLER_AUDIO_DELAY_MS <= 0:
             self._start_filler_audio(
                 epoch=epoch,