feat: improve semantic voice control and music playback

.env.example

@@ -39,6 +39,12 @@ TTS_EN_SPEAKER=en_0
 WEATHER_LAT=63.56
 WEATHER_LON=53.69
 WEATHER_CITY=Ухта
+
+# Navidrome (приоритетный источник музыки; при ошибке — fallback на Spotify)
+NAVIDROME_URL=https://navidrome.example.com
+NAVIDROME_USERNAME=your_navidrome_username
+NAVIDROME_PASSWORD=your_navidrome_password
+
 SPOTIFY_CLIENT_ID=your_spotify_client_id
 SPOTIFY_CLIENT_SECRET=your_spotify_client_secret
 SPOTIFY_REDIRECT_URI=http://localhost:8888/callback

README.md

@@ -31,7 +31,8 @@
 - Погода: текущий прогноз по городу по умолчанию или по названию города.
 - Таймеры, будильники (включая будни/выходные), секундомеры.
 - Управление громкостью системы (через `pactl`/`amixer`).
-- Управление Spotify (play/pause/next/what's playing).
+- Управление музыкой через Navidrome (приоритет) с fallback на Spotify.
+- Persistent resume: `пауза`/`продолжи` продолжают с сохранённой позиции даже после перезапуска колонки.
 - Мини-игра "Города".
 
 ## Как это работает
@@ -60,7 +61,7 @@ flowchart TD
 
 ```bash
 sudo apt-get update
-sudo apt-get install -y portaudio19-dev libasound2-dev mpg123 pulseaudio-utils alsa-utils
+sudo apt-get install -y portaudio19-dev libasound2-dev mpg123 mpv pulseaudio-utils alsa-utils
 ```
 
 ### 2) Установка Python-зависимостей
@@ -156,6 +157,9 @@ python run.py
 | `WEATHER_LAT` | Нет | - | Широта города по умолчанию |
 | `WEATHER_LON` | Нет | - | Долгота города по умолчанию |
 | `WEATHER_CITY` | Нет | `Ухта` | Город по умолчанию для погоды |
+| `NAVIDROME_URL` | Нет | - | URL Navidrome (например `https://navidrome.example.com`) |
+| `NAVIDROME_USERNAME` | Нет | - | Логин Navidrome |
+| `NAVIDROME_PASSWORD` | Нет | - | Пароль Navidrome |
 | `SPOTIFY_CLIENT_ID` | Нет | - | Spotify OAuth Client ID |
 | `SPOTIFY_CLIENT_SECRET` | Нет | - | Spotify OAuth Client Secret |
 | `SPOTIFY_REDIRECT_URI` | Нет | `http://localhost:8888/callback` | Redirect URI для Spotify |
@@ -172,7 +176,7 @@ python run.py
 | Будильник | `Поставь будильник на 7:30`, `Будильник по будням в 8:00` |
 | Секундомер | `Запусти секундомер`, `Покажи активные секундомеры` |
 | Громкость | `Громкость 7` |
-| Spotify | `Включи музыку`, `Пауза`, `Что сейчас играет` |
+| Музыка (Navidrome first) | `Включи музыку`, `Пауза`, `Продолжи`, `Следующий`, `Предыдущий`, `Что играет`, `Включи жанр electronic`, `Включи папку crystal castles` |
 | Игра | `Давай сыграем в города` |
 | Управление диалогом | `Повтори`, `Стоп`, `Хватит` |
 
@@ -222,6 +226,8 @@ alexander_smart-speaker/
 | `Audio input/output initialization failed` | проверить, что звук-сервер запущен (PipeWire/PulseAudio), и при необходимости задать `AUDIO_INPUT_DEVICE_NAME`/`AUDIO_OUTPUT_DEVICE_NAME` |
 | Будильник/таймер не звонит | наличие `mpg123` в системе |
 | Ошибка про несколько AI API | в `.env` должен остаться только один незакомментированный AI ключ |
+| Navidrome не воспроизводит | заполнены `NAVIDROME_*`, доступен `NAVIDROME_URL`, установлен `mpv` |
+| Fallback ушёл в Spotify | проверить доступность Navidrome, SSL и корректность `NAVIDROME_USERNAME`/`NAVIDROME_PASSWORD` |
 | Spotify не управляется | заполнены `SPOTIFY_*`, есть активное устройство, Premium-аккаунт |
 
 ## Лицензия

app/audio/stt.py

@@ -12,6 +12,8 @@ import re
 import time
 import pyaudio
 import logging
+import contextlib
+import threading
 from datetime import datetime, timedelta
 from ..core.config import DEEPGRAM_API_KEY, SAMPLE_RATE, WAKE_WORD_ALIASES
 from deepgram import (
@@ -29,8 +31,12 @@ from ..core.audio_manager import get_audio_manager
 _original_connect = websockets.sync.client.connect
 
 DEEPGRAM_CONNECT_TIMEOUT_SECONDS = 3.0
-DEEPGRAM_CONNECT_WAIT_SECONDS = 1.5
+DEEPGRAM_CONNECT_WAIT_SECONDS = 4.0
 DEEPGRAM_CONNECT_POLL_SECONDS = 0.001
+SENDER_STOP_WAIT_SECONDS = 2.5
+SENDER_FORCE_RELEASE_WAIT_SECONDS = 2.5
+DEEPGRAM_FINALIZATION_GRACE_SECONDS = 0.35
+DEEPGRAM_FINISH_TIMEOUT_SECONDS = 4.0
 
 
 def _patched_connect(*args, **kwargs):
@@ -152,6 +158,82 @@ class SpeechRecognizer:
                 )
         return self.stream
 
+    def _open_stream_for_session(self):
+        """Открывает отдельный входной поток для одной STT-сессии."""
+        if self.audio_manager is None:
+            self.audio_manager = get_audio_manager()
+        stream, self._input_device_index, sample_rate = self.audio_manager.open_input_stream(
+            rate=SAMPLE_RATE,
+            channels=1,
+            format=pyaudio.paInt16,
+            frames_per_buffer=4096,
+            preferred_index=self._input_device_index,
+            fallback_rates=[48000, 44100, 32000, 22050, 16000, 8000],
+        )
+        if sample_rate != SAMPLE_RATE:
+            print(
+                f"⚠️ STT mic stream uses fallback rate={sample_rate} "
+                f"(requested {SAMPLE_RATE})"
+            )
+        return stream, int(sample_rate)
+
+    def _stop_stream_quietly(self):
+        if not self.stream:
+            return
+        try:
+            if self.stream.is_active():
+                self.stream.stop_stream()
+        except Exception:
+            pass
+
+    def _release_stream(self):
+        if not self.stream:
+            return
+        self._stop_stream_quietly()
+        try:
+            self.stream.close()
+        except Exception:
+            pass
+        self.stream = None
+
+    async def _wait_for_thread(self, thread, timeout_seconds: float) -> bool:
+        """Асинхронно ждет завершения daemon-thread без блокировки event loop."""
+        deadline = time.monotonic() + timeout_seconds
+        while thread.is_alive() and time.monotonic() < deadline:
+            await asyncio.sleep(0.05)
+        return not thread.is_alive()
+
+    async def _run_blocking_cleanup(self, func, timeout_seconds: float, label: str) -> bool:
+        """Запускает потенциально подвисающий cleanup в daemon-thread и ждет ограниченное время."""
+        done_event = threading.Event()
+        error_holder = {}
+
+        def runner():
+            try:
+                func()
+            except Exception as exc:
+                error_holder["error"] = exc
+            finally:
+                done_event.set()
+
+        thread = threading.Thread(target=runner, daemon=True, name=label)
+        thread.start()
+
+        deadline = time.monotonic() + timeout_seconds
+        while not done_event.is_set() and time.monotonic() < deadline:
+            await asyncio.sleep(0.05)
+
+        if not done_event.is_set():
+            print(f"⚠️ {label} timed out; continuing cleanup.")
+            return False
+
+        error = error_holder.get("error")
+        if error is not None:
+            print(f"⚠️ {label} failed: {error}")
+            return False
+
+        return True
+
     async def _process_audio(
         self, dg_connection, timeout_seconds, detection_timeout, fast_stop
     ):
@@ -166,9 +248,9 @@ class SpeechRecognizer:
         """
         self.transcript = ""
         transcript_parts = []
+        latest_interim = ""
 
         loop = asyncio.get_running_loop()
-        stream = self._get_stream()
         effective_detection_timeout = (
             detection_timeout
             if detection_timeout is not None
@@ -192,9 +274,13 @@ class SpeechRecognizer:
         # --- Обработчики событий Deepgram ---
         def on_transcript(unused_self, result, **kwargs):
             """Вызывается, когда приходит часть текста."""
+            nonlocal latest_interim
             sentence = result.channel.alternatives[0].transcript
             if len(sentence) == 0:
                 return
+            sentence = sentence.strip()
+            if not sentence:
+                return
             try:
                 loop.call_soon_threadsafe(mark_speech_activity)
             except RuntimeError:
@@ -202,9 +288,9 @@ class SpeechRecognizer:
 
             if fast_stop:
                 if _is_fast_stop_utterance(sentence):
-                    self.transcript = sentence.strip()
+                    self.transcript = sentence
                     try:
-                        loop.call_soon_threadsafe(stop_event.set)
+                        loop.call_soon_threadsafe(request_stop)
                     except RuntimeError:
                         pass
                     return
@@ -213,6 +299,10 @@ class SpeechRecognizer:
                 # Собираем только финальные (подтвержденные) фразы
                 transcript_parts.append(sentence)
                 self.transcript = " ".join(transcript_parts).strip()
+                latest_interim = ""
+            else:
+                # Fallback: некоторые сессии завершаются без is_final.
+                latest_interim = sentence
 
         def on_speech_started(unused_self, speech_started, **kwargs):
             """Вызывается, когда VAD (Voice Activity Detection) слышит голос."""
@@ -231,7 +321,7 @@ class SpeechRecognizer:
         def on_error(unused_self, error, **kwargs):
             print(f"Deepgram Error: {error}")
             try:
-                loop.call_soon_threadsafe(stop_event.set)
+                loop.call_soon_threadsafe(request_stop)
             except RuntimeError:
                 # Event loop might be closed, ignore
                 pass
@@ -242,27 +332,34 @@ class SpeechRecognizer:
         dg_connection.on(LiveTranscriptionEvents.UtteranceEnd, on_utterance_end)
         dg_connection.on(LiveTranscriptionEvents.Error, on_error)
 
-        # Параметры распознавания
-        options = LiveOptions(
-            model="nova-2",  # Самая быстрая и точная модель
-            language=self.current_lang,
-            smart_format=True,  # Расстановка знаков препинания
-            encoding="linear16",
-            channels=1,
-            sample_rate=self._stream_sample_rate,
-            interim_results=True,
-            utterance_end_ms=int(POST_SPEECH_SILENCE_TIMEOUT_SECONDS * 1000),
-            vad_events=True,
-            # Сглаженный порог endpointing, чтобы не резать речь на коротких паузах.
-            endpointing=int(POST_SPEECH_SILENCE_TIMEOUT_SECONDS * 1000),
-        )
-
         # --- Задача отправки аудио с буферизацией ---
-        async def send_audio():
+        sender_stop_event = threading.Event()
+
+        def request_stop():
+            stop_event.set()
+            sender_stop_event.set()
+
+        def send_audio():
             chunks_sent = 0
             audio_buffer = []  # Буфер для накопления звука во время подключения
+            stream = None
 
             try:
+                stream, stream_sample_rate = self._open_stream_for_session()
+                options = LiveOptions(
+                    model="nova-2",  # Самая быстрая и точная модель
+                    language=self.current_lang,
+                    smart_format=True,  # Расстановка знаков препинания
+                    encoding="linear16",
+                    channels=1,
+                    sample_rate=stream_sample_rate,
+                    interim_results=True,
+                    utterance_end_ms=int(POST_SPEECH_SILENCE_TIMEOUT_SECONDS * 1000),
+                    vad_events=True,
+                    # Сглаженный порог endpointing, чтобы не резать речь на коротких паузах.
+                    endpointing=int(POST_SPEECH_SILENCE_TIMEOUT_SECONDS * 1000),
+                )
+
                 # 1. Сразу начинаем захват звука, не дожидаясь сети!
                 stream.start_stream()
                 print("🎤 Stream started (buffering)...")
@@ -270,34 +367,61 @@ class SpeechRecognizer:
                 # 2. Запускаем подключение к Deepgram в фоне (через ThreadPool, т.к. start() блокирующий)
                 # Но в данном SDK start() возвращает bool, он может быть блокирующим.
                 # Deepgram Python SDK v3+ start() делает handshake.
+                connect_result = {"done": False, "ok": None, "error": None}
 
-                connect_future = loop.run_in_executor(
+                def start_connection():
-                    None, lambda: dg_connection.start(options)
+                    try:
+                        connect_result["ok"] = dg_connection.start(options)
+                    except Exception as exc:
+                        connect_result["error"] = exc
+                    finally:
+                        connect_result["done"] = True
+
+                connect_thread = threading.Thread(
+                    target=start_connection, daemon=True
                 )
+                connect_thread.start()
 
                 # Пока подключаемся, копим данные.
                 # Ждём коротко: если сеть подвисла, быстрее перезапускаем попытку.
                 connect_deadline = time.monotonic() + DEEPGRAM_CONNECT_WAIT_SECONDS
                 while (
-                    not connect_future.done()
+                    not connect_result["done"]
                     and time.monotonic() < connect_deadline
+                    and not sender_stop_event.is_set()
                 ):
                     if stream.is_active():
-                        data = stream.read(4096, exception_on_overflow=False)
+                        try:
+                            data = stream.read(4096, exception_on_overflow=False)
+                        except Exception as read_error:
+                            if sender_stop_event.is_set():
+                                return
+                            print(f"Audio read error during connect: {read_error}")
+                            with contextlib.suppress(RuntimeError):
+                                loop.call_soon_threadsafe(request_stop)
+                            return
                         audio_buffer.append(data)
-                    await asyncio.sleep(DEEPGRAM_CONNECT_POLL_SECONDS)
+                    time.sleep(DEEPGRAM_CONNECT_POLL_SECONDS)
 
-                if not connect_future.done():
+                if sender_stop_event.is_set():
+                    return
+
+                if not connect_result["done"]:
                     print(
                         f"⏰ Timeout connecting to Deepgram ({DEEPGRAM_CONNECT_WAIT_SECONDS:.1f}s)"
                     )
-                    stop_event.set()
+                    loop.call_soon_threadsafe(request_stop)
                     return
 
                 # Проверяем результат подключения
-                if connect_future.result() is False:
+                if connect_result["error"] is not None:
+                    print(f"Failed to start Deepgram connection: {connect_result['error']}")
+                    loop.call_soon_threadsafe(request_stop)
+                    return
+
+                if connect_result["ok"] is False:
                     print("Failed to start Deepgram connection")
-                    stop_event.set()
+                    loop.call_soon_threadsafe(request_stop)
                     return
 
                 print(f"🚀 Connected! Sending buffer ({len(audio_buffer)} chunks)...")
@@ -310,23 +434,45 @@ class SpeechRecognizer:
                 audio_buffer = None  # Освобождаем память
 
                 # 4. Продолжаем стримить в реальном времени до события остановки.
-                while not stop_event.is_set():
+                while not sender_stop_event.is_set():
-                    if stream.is_active():
+                    if not stream.is_active():
+                        break
+                    try:
                         data = stream.read(4096, exception_on_overflow=False)
-                        dg_connection.send(data)
+                    except Exception as read_error:
-                        chunks_sent += 1
+                        if sender_stop_event.is_set():
-                        if chunks_sent % 50 == 0:
+                            break
-                            print(".", end="", flush=True)
+                        print(f"Audio read error: {read_error}")
-                    await asyncio.sleep(0.002)  # Уменьшаем задержку для более быстрого реагирования
+                        with contextlib.suppress(RuntimeError):
+                            loop.call_soon_threadsafe(request_stop)
+                        break
+                    if sender_stop_event.is_set():
+                        break
+                    dg_connection.send(data)
+                    chunks_sent += 1
+                    if chunks_sent % 50 == 0:
+                        print(".", end="", flush=True)
+                    time.sleep(0.002)  # Уменьшаем задержку для более быстрого реагирования
 
             except Exception as e:
                 print(f"Audio send error: {e}")
+                with contextlib.suppress(RuntimeError):
+                    loop.call_soon_threadsafe(request_stop)
             finally:
-                if stream.is_active():
+                with contextlib.suppress(Exception):
-                    stream.stop_stream()
+                    if stream and stream.is_active():
+                        stream.stop_stream()
+                with contextlib.suppress(Exception):
+                    if stream:
+                        stream.close()
                 print(f"\n🛑 Stream stopped. Chunks sent: {chunks_sent}")
 
-        sender_task = asyncio.create_task(send_audio())
+        sender_thread = threading.Thread(
+            target=send_audio,
+            daemon=True,
+            name="deepgram-audio-sender",
+        )
+        sender_thread.start()
 
         if False:  # dg_connection.start(options) перенесен внутрь send_audio
             pass
@@ -356,7 +502,7 @@ class SpeechRecognizer:
 
                 if not done:
                     # Если за detection_timeout никто не начал говорить, выходим
-                    stop_event.set()
+                    request_stop()
 
             # 2. После старта речи завершаем только по тишине POST_SPEECH_SILENCE_TIMEOUT_SECONDS.
             # Добавляем длинный защитный лимит, чтобы сессия не зависла навсегда.
@@ -374,7 +520,7 @@ class SpeechRecognizer:
                             now - last_speech_activity
                             >= POST_SPEECH_SILENCE_TIMEOUT_SECONDS
                         ):
-                            stop_event.set()
+                            request_stop()
                             break
 
                         if (
@@ -383,7 +529,7 @@ class SpeechRecognizer:
                             >= max_active_speech_seconds
                         ):
                             print("⏱️ Достигнут защитный лимит активного прослушивания.")
-                            stop_event.set()
+                            request_stop()
                             break
 
                     await asyncio.sleep(0.05)
@@ -393,19 +539,29 @@ class SpeechRecognizer:
         except Exception as e:
             print(f"Error in waiting for events: {e}")
 
-        stop_event.set()
+        request_stop()
-        try:
+        sender_stopped = await self._wait_for_thread(
-            await sender_task
+            sender_thread,
-        except Exception as e:
+            timeout_seconds=max(SENDER_STOP_WAIT_SECONDS, SENDER_FORCE_RELEASE_WAIT_SECONDS),
-            print(f"Error waiting for sender task: {e}")
+        )
+        if not sender_stopped:
+            print("⚠️ Audio sender shutdown timed out; continuing cleanup.")
+
+        # Небольшая пауза, чтобы получить последние transcript-события перед finish().
+        await asyncio.sleep(DEEPGRAM_FINALIZATION_GRACE_SECONDS)
 
         # Завершаем соединение и ждем последние результаты
-        try:
+        await self._run_blocking_cleanup(
-            dg_connection.finish()
+            dg_connection.finish,
-        except Exception as e:
+            timeout_seconds=DEEPGRAM_FINISH_TIMEOUT_SECONDS,
-            print(f"Error finishing connection: {e}")
+            label="Deepgram finish",
+        )
 
-        return self.transcript
+        final_text = self.transcript.strip()
+        if not final_text:
+            final_text = latest_interim.strip()
+        self.transcript = final_text
+        return final_text
 
     def listen(
         self,

app/audio/tts.py

@@ -286,6 +286,9 @@ class TextToSpeech:
         if not text.strip():
             return True
 
+        if check_interrupt is None:
+            check_interrupt = self._default_interrupt_checker()
+
         if language == "ru":
             text = self._preprocess_text(text)
             segments = self._split_mixed_language(text)
@@ -296,6 +299,14 @@ class TextToSpeech:
             text, check_interrupt=check_interrupt, language=language
         )
 
+    def _default_interrupt_checker(self):
+        try:
+            from .wakeword import check_wakeword_once
+
+            return check_wakeword_once
+        except Exception:
+            return None
+
     def _resample_audio(self, audio_np: np.ndarray, src_rate: int, dst_rate: int):
         if src_rate == dst_rate:
             return audio_np.astype(np.float32, copy=False)

app/core/ai.py

@@ -54,6 +54,26 @@ No explanations, no quotes, no comments.
 Separate variants with " / " (space slash space).
 Keep the translation максимально кратким и естественным, без лишних слов."""
 
+INTENT_SYSTEM_PROMPT = """Ты NLU-модуль голосовой колонки.
+Твоя задача: распознать намерение пользователя и вернуть СТРОГО JSON без markdown и пояснений.
+Всегда возвращай объект c ключами:
+{
+  "intent": "none|music|timer|alarm|weather|volume|translation|cities|repeat|stop|smalltalk|chat",
+  "normalized_command": "<краткая нормализованная команда на русском или пусто>",
+  "music_action": "none|play|pause|resume|next|previous|current|play_genre|play_folder|play_query",
+  "music_query": "<запрос для музыки/жанра/папки или пусто>",
+  "confidence": 0.0
+}
+Правила:
+- Если это музыка, ставь intent=music и выбирай music_action.
+- "Включи музыку" и любые эквиваленты = music_action=play.
+- Для "пауза/останови музыку/выключи музыку" = music_action=pause.
+- Для "что играет" = music_action=current.
+- Для "включи жанр X" = music_action=play_genre, music_query=X.
+- Для "включи папку X" = music_action=play_folder, music_query=X.
+- normalized_command должен быть пригоден для командного парсера (без лишних слов).
+- Если уверенность низкая, ставь intent=none, music_action=none, confidence <= 0.4."""
+
 _PROVIDER_ALIASES = {
     "": "openrouter",
     "anthropic": "anthropic",
@@ -381,6 +401,32 @@ def _log_request_exception(cfg, error: Exception):
     print(f"❌ Ошибка API ({cfg['name']}): {error}{details}")
 
 
+def _extract_json_object(raw_text: str) -> Optional[dict]:
+    text = str(raw_text or "").strip()
+    if not text:
+        return None
+
+    try:
+        payload = json.loads(text)
+        if isinstance(payload, dict):
+            return payload
+    except json.JSONDecodeError:
+        pass
+
+    match = re.search(r"\{.*\}", text, flags=re.DOTALL)
+    if not match:
+        return None
+
+    candidate = match.group(0).strip()
+    try:
+        payload = json.loads(candidate)
+    except json.JSONDecodeError:
+        return None
+    if isinstance(payload, dict):
+        return payload
+    return None
+
+
 def _send_request(messages, max_tokens, temperature, error_text):
     """
     Внутренняя функция для отправки HTTP-запроса к выбранному AI-провайдеру.
@@ -422,6 +468,98 @@ def _send_request(messages, max_tokens, temperature, error_text):
         return "Не удалось обработать ответ от AI."
 
 
+def interpret_assistant_intent(text: str) -> dict:
+    """
+    Interprets voice command semantics for downstream command routers.
+    Returns a normalized dict even when AI is unavailable.
+    """
+    result = {
+        "intent": "none",
+        "normalized_command": "",
+        "music_action": "none",
+        "music_query": "",
+        "confidence": 0.0,
+    }
+    cleaned_text = str(text or "").strip()
+    if not cleaned_text:
+        return result
+
+    cfg, selection_error = _get_provider_settings()
+    if selection_error:
+        return result
+    if _get_provider_config_error(cfg):
+        return result
+
+    messages = [
+        {"role": "system", "content": INTENT_SYSTEM_PROMPT},
+        {"role": "user", "content": cleaned_text},
+    ]
+    response = _send_request(
+        messages,
+        max_tokens=220,
+        temperature=0.0,
+        error_text="",
+    )
+    payload = _extract_json_object(response)
+    if not payload:
+        return result
+
+    allowed_intents = {
+        "none",
+        "music",
+        "timer",
+        "alarm",
+        "weather",
+        "volume",
+        "translation",
+        "cities",
+        "repeat",
+        "stop",
+        "smalltalk",
+        "chat",
+    }
+    allowed_music_actions = {
+        "none",
+        "play",
+        "pause",
+        "resume",
+        "next",
+        "previous",
+        "current",
+        "play_genre",
+        "play_folder",
+        "play_query",
+    }
+
+    intent = str(payload.get("intent", "none")).strip().lower()
+    if intent not in allowed_intents:
+        intent = "none"
+
+    music_action = str(payload.get("music_action", "none")).strip().lower()
+    if music_action not in allowed_music_actions:
+        music_action = "none"
+
+    try:
+        confidence = float(payload.get("confidence", 0.0))
+    except (TypeError, ValueError):
+        confidence = 0.0
+    confidence = max(0.0, min(1.0, confidence))
+
+    normalized_command = str(payload.get("normalized_command", "")).strip()
+    music_query = str(payload.get("music_query", "")).strip()
+
+    result.update(
+        {
+            "intent": intent,
+            "normalized_command": normalized_command,
+            "music_action": music_action,
+            "music_query": music_query,
+            "confidence": confidence,
+        }
+    )
+    return result
+
+
 def ask_ai(messages_history: list) -> str:
     """
     Запрос к AI в режиме чата.

app/core/config.py

@@ -129,3 +129,8 @@ TTS_SAMPLE_RATE = 48000
 WEATHER_LAT = os.getenv("WEATHER_LAT")
 WEATHER_LON = os.getenv("WEATHER_LON")
 WEATHER_CITY = os.getenv("WEATHER_CITY", "Ухта")
+
+# --- Настройки Navidrome (музыка) ---
+NAVIDROME_URL = os.getenv("NAVIDROME_URL", "").strip().rstrip("/")
+NAVIDROME_USERNAME = os.getenv("NAVIDROME_USERNAME", "").strip()
+NAVIDROME_PASSWORD = os.getenv("NAVIDROME_PASSWORD", "")

app/main.py

@@ -33,7 +33,7 @@ from .audio.wakeword import (
 from .audio.wakeword import (
     stop_monitoring as stop_wakeword_monitoring,
 )
-from .core.ai import ask_ai_stream, translate_text
+from .core.ai import ask_ai_stream, interpret_assistant_intent, translate_text
 from .core.config import BASE_DIR, WAKE_WORD
 from .core.cleaner import clean_response
 from .core.commands import is_stop_command
@@ -163,6 +163,10 @@ _CITY_PATTERNS = [
     ),
 ]
 
+_SEMANTIC_INTENT_MIN_CONFIDENCE = 0.55
+_SEMANTIC_MUSIC_MIN_CONFIDENCE = 0.45
+_SEMANTIC_REPEAT_STOP_MIN_CONFIDENCE = 0.72
+
 
 def signal_handler(sig, frame):
     """Обработчик Ctrl+C."""
@@ -311,7 +315,7 @@ def main():
                     continue  # Продолжаем цикл
             else:
                 # Follow-up режим — без wake word
-                print(f"👂 Слушаю ({followup_idle_timeout_seconds:.0f} сек)...")
+                print(f"👂 Слушаю ({followup_idle_timeout_seconds:.1f} сек)...")
                 try:
                     user_text = listen(
                         timeout_seconds=7.0,
@@ -341,6 +345,11 @@ def main():
 
             # Проверка на команду "Стоп"
             if is_stop_command(user_text):
+                music_controller = get_music_controller()
+                music_stop_response = music_controller.pause_for_stop_word()
+                if music_stop_response:
+                    print(f"🎵 {music_stop_response}")
+
                 if stopwatch_manager.has_running_stopwatches():
                     stopwatch_stop_response = stopwatch_manager.pause_stopwatches()
                     clean_stopwatch_stop_response = clean_response(
@@ -369,8 +378,93 @@ def main():
                 skip_wakeword = True
                 continue
 
+            effective_text = user_text
+            semantic_intent = interpret_assistant_intent(user_text)
+            semantic_type = str(semantic_intent.get("intent", "none")).strip().lower()
+            try:
+                semantic_confidence = float(
+                    semantic_intent.get("confidence", 0.0) or 0.0
+                )
+            except (TypeError, ValueError):
+                semantic_confidence = 0.0
+            semantic_command = str(semantic_intent.get("normalized_command", "")).strip()
+            semantic_music_action = (
+                str(semantic_intent.get("music_action", "none")).strip().lower()
+            )
+            semantic_music_query = str(semantic_intent.get("music_query", "")).strip()
+
+            if (
+                semantic_type == "stop"
+                and semantic_confidence >= _SEMANTIC_REPEAT_STOP_MIN_CONFIDENCE
+            ):
+                music_controller = get_music_controller()
+                music_stop_response = music_controller.pause_for_stop_word()
+                if music_stop_response:
+                    print(f"🎵 {music_stop_response}")
+
+                if stopwatch_manager.has_running_stopwatches():
+                    stopwatch_stop_response = stopwatch_manager.pause_stopwatches()
+                    clean_stopwatch_stop_response = clean_response(
+                        stopwatch_stop_response, language="ru"
+                    )
+                    speak(clean_stopwatch_stop_response)
+                    last_response = clean_stopwatch_stop_response
+                    skip_wakeword = False
+                    continue
+                print("_" * 50)
+                print(f"💤 Жду '{WAKE_WORD}'...")
+                skip_wakeword = False
+                continue
+
+            if (
+                semantic_type == "repeat"
+                and semantic_confidence >= _SEMANTIC_REPEAT_STOP_MIN_CONFIDENCE
+            ):
+                if last_response:
+                    print(f"🔁 Повторяю: {last_response}")
+                    speak(last_response)
+                else:
+                    speak("Я еще ничего не говорил.")
+                skip_wakeword = True
+                continue
+
+            if (
+                semantic_type == "music"
+                and semantic_confidence >= _SEMANTIC_MUSIC_MIN_CONFIDENCE
+            ):
+                music_controller = get_music_controller()
+                semantic_music_response = music_controller.handle_semantic_action(
+                    semantic_music_action,
+                    semantic_music_query,
+                )
+                if semantic_music_response:
+                    clean_music_response = clean_response(
+                        semantic_music_response, language="ru"
+                    )
+                    speak(clean_music_response)
+                    last_response = clean_music_response
+                    skip_wakeword = True
+                    continue
+
+            if (
+                semantic_command
+                and semantic_confidence >= _SEMANTIC_INTENT_MIN_CONFIDENCE
+                and semantic_type
+                in {
+                    "music",
+                    "timer",
+                    "alarm",
+                    "weather",
+                    "volume",
+                    "translation",
+                    "cities",
+                }
+            ):
+                effective_text = semantic_command
+                print(f"🧠 Команда: '{user_text}' -> '{effective_text}'")
+
             # Small-talk
-            smalltalk_response = get_smalltalk_response(user_text)
+            smalltalk_response = get_smalltalk_response(effective_text)
             if smalltalk_response:
                 clean_smalltalk = clean_response(smalltalk_response, language="ru")
                 speak(clean_smalltalk)
@@ -378,7 +472,7 @@ def main():
                 skip_wakeword = True
                 continue
 
-            command_text = user_text
+            command_text = effective_text
             command_text_lower = command_text.lower()
             if pending_time_target == "timer" and "таймер" not in command_text_lower:
                 command_text = f"таймер {command_text}"
@@ -427,9 +521,9 @@ def main():
                 continue
 
             # Громкость
-            if user_text.lower().startswith("громкость"):
+            if command_text.lower().startswith("громкость"):
                 try:
-                    vol_str = user_text.lower().replace("громкость", "", 1).strip()
+                    vol_str = command_text.lower().replace("громкость", "", 1).strip()
                     level = parse_volume_text(vol_str)
 
                     if level is not None:
@@ -455,7 +549,7 @@ def main():
 
             # Погода
             requested_city = None
-            user_text_lower = user_text.lower()
+            user_text_lower = command_text.lower()
 
             for pattern in _CITY_PATTERNS:
                 match = pattern.search(user_text_lower)
@@ -487,7 +581,7 @@ def main():
 
             # Музыка
             music_controller = get_music_controller()
-            music_response = music_controller.parse_command(user_text)
+            music_response = music_controller.parse_command(command_text)
             if music_response:
                 clean_music_response = clean_response(music_response, language="ru")
                 speak(clean_music_response)
@@ -496,7 +590,7 @@ def main():
                 continue
 
             # Перевод
-            translation_request = parse_translation_request(user_text)
+            translation_request = parse_translation_request(command_text)
             if translation_request:
                 source_lang = translation_request["source_lang"]
                 target_lang = translation_request["target_lang"]
@@ -553,8 +647,7 @@ def main():
                 continue
 
             # Игра "Города"
-            cities_response = cities_game.handle(user_text)
+            cities_response = cities_game.handle(command_text)
-            cities_response = cities_game.handle(user_text)
             if cities_response:
                 clean_cities_response = clean_response(cities_response, language="ru")
                 speak(clean_cities_response)