Trixy7/trixy_core/wakeword/service.py

# -*- coding: utf-8 -*-
"""
Wakeword Service - Haupt-Service für Wakeword-Erkennung.

Läuft auf Satellite und Standalone.

Audio-Verarbeitung läuft in einem dedizierten Thread (nicht im asyncio-Loop),
analog zu den OpenWakeWord-Beispielen. Nur Netzwerk-I/O und Event-Emission
werden auf den asyncio-Loop dispatcht.
"""

from dataclasses import dataclass, field
from datetime import datetime
from enum import Enum
from typing import TYPE_CHECKING, Callable, Any, Awaitable
import asyncio
import queue
import time
import threading

from trixy_core.service import IService, ServicePriority, ServiceGroup
from trixy_core.wakeword.detector import (
    WakewordDetector,
    WakewordDetection,
    DetectorConfig,
    WakewordType,
)
from trixy_core.wakeword.audio_buffer import AudioBuffer, BufferConfig
from trixy_core.wakeword.vad import VoiceActivityDetector, VADConfig, VADState
from trixy_core.utils.debug import pinfo, pdebug, perror, pwarn

if TYPE_CHECKING:
    from trixy_core.application import BaseApplication
    from trixy_core.audio.microphone import MicrophoneCapture
    from trixy_core.network.client_connection import ClientConnection


class ServiceState(Enum):
    """Zustand des Wakeword-Service."""
    STOPPED = "stopped"           # Service gestoppt
    LISTENING = "listening"       # Wakeword-Erkennung aktiv
    PAUSED = "paused"             # Pausiert (z.B. während Arbitration)
    RECORDING = "recording"       # Aufnahme läuft
    WAITING_RESPONSE = "waiting"  # Warte auf Antwort vom Server
    PROCESSING = "processing"     # Verarbeitung läuft
    FOLLOW_UP = "follow_up"       # Warte auf Rückfrage-Antwort


@dataclass
class WakewordServiceConfig:
    """Konfiguration für den Wakeword-Service."""

    # Detector-Konfiguration
    detector: DetectorConfig = field(default_factory=DetectorConfig)

    # Buffer-Konfiguration
    buffer: BufferConfig = field(default_factory=BufferConfig)

    # VAD-Konfiguration
    vad: VADConfig = field(default_factory=VADConfig)

    # Aufnahme-Konfiguration
    max_recording_seconds: float = 60.0
    silence_timeout_seconds: float = 3.0
    follow_up_timeout_seconds: float = 60.0

    # Arbitration
    arbitration_timeout_seconds: float = 2.0

    # Cooldown nach Wakeword-Erkennung (verhindert Mehrfach-Erkennung)
    wakeword_cooldown_seconds: float = 3.0

    # Standalone-Modus (keine Arbitration)
    standalone_mode: bool = False

    # Audio-Quelle
    audio_source: str | None = None  # None = Standard-Mikrofon


@dataclass
class RecordingSession:
    """Eine Aufnahme-Session."""

    session_id: str
    wakeword_detection: WakewordDetection
    start_time: datetime
    end_time: datetime | None = None
    audio_data: bytes = b""
    state: ServiceState = ServiceState.RECORDING
    is_selected: bool = False

    # Für Rückfragen
    follow_up_count: int = 0
    last_activity: datetime = field(default_factory=datetime.now)


class WakewordService(IService):
    """
    Wakeword Detection Service.

    Verantwortlich für:
    - Wakeword-Erkennung in Echtzeit
    - Audio-Aufnahme nach Wakeword
    - VAD für Sprach-Ende-Erkennung
    - Kommunikation mit Server (Arbitration)

    Audio-Verarbeitung läuft in einem dedizierten Thread, um den
    asyncio-Event-Loop nicht zu blockieren (wie in den OpenWakeWord-Beispielen).
    """

    PRIORITY = ServicePriority.CORE
    GROUP = ServiceGroup.AUDIO
    DEPENDENCIES = []

    def __init__(
        self,
        application: "BaseApplication",
        config: WakewordServiceConfig | None = None,
    ):
        """
        Initialisiert den Wakeword-Service.

        Args:
            application: Anwendungs-Instanz
            config: Service-Konfiguration
        """
        super().__init__(application)

        self._config = config or WakewordServiceConfig()

        # Komponenten
        self._detector = WakewordDetector(self._config.detector)
        self._audio_buffer = AudioBuffer(self._config.buffer)
        self._vad = VoiceActivityDetector(self._config.vad)

        # State
        self._state = ServiceState.STOPPED
        self._state_lock = threading.RLock()
        self._current_session: RecordingSession | None = None

        # Audio-Thread (dedizierter Thread für Echtzeit-Verarbeitung)
        self._audio_thread: threading.Thread | None = None
        self._audio_queue: queue.Queue[bytes] = queue.Queue(maxsize=50)
        self._stop_event = threading.Event()

        # Referenz auf den asyncio-Event-Loop (für Thread → Loop Kommunikation)
        self._loop: asyncio.AbstractEventLoop | None = None

        # Callbacks
        self._on_wakeword: list[Callable[[WakewordDetection], None]] = []
        self._on_recording_complete: list[Callable[[RecordingSession], None]] = []
        self._on_state_change: list[Callable[[ServiceState], None]] = []

        # Server-Kommunikation (wird von Client gesetzt)
        self._send_to_server: Callable[[str, dict], asyncio.Future] | None = None
        self._session_id_generator: Callable[[], str] | None = None

        # Mikrofon (Client-Modus)
        self._microphone: MicrophoneCapture | None = None

        # Audio-Streaming an Server (connection.send_audio)
        self._audio_streamer: Callable[[bytes], Awaitable[bool]] | None = None

        # Pending Future für Arbitration-Antwort
        self._arbitration_future: asyncio.Future | None = None

        # Cooldown für Wakeword-Erkennung (verhindert Mehrfach-Erkennung)
        self._last_wakeword_time: float = 0.0

        # Audio-Ducking: ALSA-Lautstaerke bei Wakeword reduzieren (Client-seitig)
        self._volume_handler: object | None = None  # VolumeHandler Instanz
        self._ducking_volume_pct: int = 10           # Ziel-Lautstaerke beim Ducken
        self._pre_duck_volume: int | None = None     # Gespeicherte Original-Lautstaerke
        self._is_ducked: bool = False

        # Externe Pause (Server/Plugin-gesteuert)
        self._externally_paused: bool = False

        # Server-gesteuerte Aufnahme
        self._server_recording_active: bool = False
        self._server_recording_id: str = ""
        self._server_recording_start_time: float = 0.0
        self._on_server_recording_stopped: Callable[[str, float], Awaitable[None]] | None = None

    @property
    def config(self) -> WakewordServiceConfig:
        """Gibt Konfiguration zurück."""
        return self._config

    @property
    def state(self) -> ServiceState:
        """Aktueller Zustand."""
        with self._state_lock:
            return self._state

    @property
    def externally_paused(self) -> bool:
        """Prüft ob die Erkennung extern pausiert ist."""
        return self._externally_paused

    @property
    def is_listening(self) -> bool:
        """Prüft ob Service auf Wakeword hört."""
        return self._state == ServiceState.LISTENING

    @property
    def is_recording(self) -> bool:
        """Prüft ob Aufnahme läuft."""
        return self._state in (ServiceState.RECORDING, ServiceState.FOLLOW_UP)

    @property
    def current_session(self) -> RecordingSession | None:
        """Aktuelle Aufnahme-Session."""
        return self._current_session

    def _set_state(self, new_state: ServiceState) -> None:
        """Setzt neuen Zustand und triggert Callbacks + Events."""
        with self._state_lock:
            if self._state == new_state:
                return

            old_state = self._state
            self._state = new_state

        # Callbacks außerhalb des Locks
        for callback in self._on_state_change:
            try:
                callback(new_state)
            except Exception:
                pass

        pdebug(f"WakewordService State: {old_state.value} -> {new_state.value}")

        # Events fuer wichtige State-Wechsel (async auf Event-Loop dispatchen)
        event_map = {
            ServiceState.LISTENING: "wakeword_listening",
            ServiceState.PAUSED: "wakeword_paused",
            ServiceState.RECORDING: "recording_started",
            ServiceState.PROCESSING: "recording_stopped",
        }
        event_name = event_map.get(new_state)
        if event_name:
            event_data = {"old_state": old_state.value, "new_state": new_state.value}
            if self._current_session:
                event_data["session_id"] = self._current_session.session_id
            if event_name == "recording_stopped" and self._current_session:
                elapsed = (datetime.now() - self._current_session.start_time).total_seconds()
                event_data["duration_seconds"] = elapsed
            self._emit_event(event_name, event_data)

    # === Audio-Ducking (ALSA, laeuft im Audio-Thread) ===

    def set_volume_handler(self, handler: object, ducking_pct: int = 10) -> None:
        """
        Setzt den VolumeHandler fuer Audio-Ducking.

        Wird vom Client bei der Initialisierung aufgerufen.

        Args:
            handler: VolumeHandler Instanz (hat get_volume/set_volume)
            ducking_pct: Ziel-Lautstaerke beim Ducken (0-100)
        """
        self._volume_handler = handler
        self._ducking_volume_pct = ducking_pct
        pdebug(f"WakewordService: Audio-Ducking aktiviert ({ducking_pct}%)")

    def _duck_audio(self) -> None:
        """
        Reduziert die ALSA-Lautstaerke sofort (synchron, im Audio-Thread).

        Nutzt amixer direkt statt den async VolumeHandler,
        weil wir im Audio-Thread sind und nicht warten koennen.
        """
        if self._is_ducked:
            return

        import subprocess
        import shutil

        if not shutil.which("amixer"):
            return

        try:
            # Aktuelle Lautstaerke lesen
            result = subprocess.run(
                ["amixer", "get", "Master"],
                capture_output=True, text=True, timeout=2,
            )
            if result.returncode == 0:
                for line in result.stdout.splitlines():
                    if "%" in line:
                        start = line.index("[") + 1
                        end = line.index("%")
                        self._pre_duck_volume = int(line[start:end])
                        break

            # Lautstaerke reduzieren
            subprocess.run(
                ["amixer", "set", "Master", f"{self._ducking_volume_pct}%"],
                capture_output=True, timeout=2,
            )
            self._is_ducked = True
            pdebug(f"Audio-Duck: {self._pre_duck_volume}% → {self._ducking_volume_pct}%")

        except Exception as e:
            pdebug(f"Audio-Duck Fehler: {e}")

    def _unduck_audio(self) -> None:
        """Stellt die ALSA-Lautstaerke wieder her (synchron)."""
        if not self._is_ducked or self._pre_duck_volume is None:
            return

        import subprocess
        import shutil

        if not shutil.which("amixer"):
            return

        try:
            subprocess.run(
                ["amixer", "set", "Master", f"{self._pre_duck_volume}%"],
                capture_output=True, timeout=2,
            )
            pdebug(f"Audio-Unduck: {self._ducking_volume_pct}% → {self._pre_duck_volume}%")
        except Exception as e:
            pdebug(f"Audio-Unduck Fehler: {e}")

        self._pre_duck_volume = None
        self._is_ducked = False

    def _emit_event(self, event_name: str, data: dict) -> None:
        """Emittiert ein Event thread-safe auf dem asyncio-Loop."""
        events = getattr(self._application, "events", None)
        if events and self._loop and self._loop.is_running():
            self._schedule_async(events.emit(event_name, data))

    def _schedule_async(self, coro: Any) -> None:
        """
        Plant eine Coroutine thread-safe auf dem asyncio-Loop ein.

        Wird aus dem Audio-Thread aufgerufen für Netzwerk-I/O und Events.
        """
        if self._loop and self._loop.is_running():
            asyncio.run_coroutine_threadsafe(coro, self._loop)

    async def start(self) -> None:
        """Startet den Service (IService)."""
        await self.on_start()

    async def stop(self) -> None:
        """Stoppt den Service (IService)."""
        await self.on_stop()

    async def on_start(self) -> None:
        """Service-Start."""
        pinfo("Lade Wakeword-Modelle...")

        try:
            self._detector.load_models()
            pinfo(f"Wakeword-Modelle geladen: {self._detector.loaded_models}")
        except Exception as e:
            perror(f"Fehler beim Laden der Wakeword-Modelle: {e}")
            raise

        # Originale Aufnahme-Konfiguration merken (fuer recording_config_override Reset)
        self._original_silence_timeout = self._config.silence_timeout_seconds
        self._original_max_duration_ms = self._config.vad.max_duration_ms

        # Event-Handler fuer temporaere Konfigurationsaenderungen registrieren
        event_manager = getattr(self._application, "events", None)
        if event_manager:
            event_manager.register("recording_config_override", self._on_config_override)
            event_manager.register("wakeword_manual_trigger", self._on_manual_trigger)

        # Registriere Detector-Callback
        self._detector.on_detection(self._on_wakeword_detected)

        # VAD-Callbacks
        self._vad.on_silence(self._on_silence_detected)
        self._vad.on_timeout(self._on_recording_timeout)
        self._vad.on_no_speech(self._on_no_speech_detected)

    async def on_stop(self) -> None:
        """Service-Stop."""
        await self.stop_listening()
        self._detector.unload_models()

    async def start_listening(self) -> None:
        """Startet die Wakeword-Erkennung."""
        if self._state != ServiceState.STOPPED:
            return

        # Event-Loop-Referenz speichern
        self._loop = asyncio.get_running_loop()

        # Queue leeren (alte Frames von vorherigem Lauf verwerfen)
        self._drain_audio_queue()

        self._set_state(ServiceState.LISTENING)
        self._stop_event.clear()

        # Starte dedizierten Audio-Thread
        self._audio_thread = threading.Thread(
            target=self._audio_processing_loop,
            name="wakeword-audio",
            daemon=True,
        )
        self._audio_thread.start()

        # Mikrofon starten und direkt mit Queue verbinden
        if self._microphone:
            self._microphone.set_callback(self._mic_callback)
            self._microphone.start()

        pinfo("Wakeword-Erkennung gestartet")

    async def stop_listening(self) -> None:
        """Stoppt die Wakeword-Erkennung."""
        if self._microphone:
            self._microphone.stop()

        self._stop_event.set()

        if self._audio_thread and self._audio_thread.is_alive():
            self._audio_thread.join(timeout=2.0)
            self._audio_thread = None

        self._set_state(ServiceState.STOPPED)
        pinfo("Wakeword-Erkennung gestoppt")

    def _mic_callback(self, data: bytes) -> None:
        """
        Mikrofon-Callback (läuft in PortAudio-Thread).

        Ringbuffer-Semantik: Bei voller Queue wird der älteste Frame verworfen.
        Kein call_soon_threadsafe nötig — queue.Queue ist thread-safe.
        """
        try:
            self._audio_queue.put_nowait(data)
        except queue.Full:
            # Queue voll: ältesten Frame verwerfen, neuen einfügen
            try:
                self._audio_queue.get_nowait()
            except queue.Empty:
                pass
            try:
                self._audio_queue.put_nowait(data)
            except queue.Full:
                pass

    def pause(self) -> None:
        """Pausiert die Wakeword-Erkennung."""
        if self._state == ServiceState.LISTENING:
            self._set_state(ServiceState.PAUSED)

    def _drain_audio_queue(self) -> None:
        """Leert die Audio-Queue (verwirft aufgestaute Frames)."""
        drained = 0
        while True:
            try:
                self._audio_queue.get_nowait()
                drained += 1
            except queue.Empty:
                break
        if drained > 0:
            pdebug(f"Audio-Queue geleert: {drained} Frames verworfen")

    def resume(self) -> None:
        """Setzt die Wakeword-Erkennung nach interner Verarbeitung fort."""
        if self._state == ServiceState.STOPPED:
            return

        # Audio-Ducking: Lautstaerke wiederherstellen
        self._unduck_audio()

        # Queue leeren (alle Frames aus der Recording-Phase verwerfen)
        self._drain_audio_queue()
        self._detector.reset()

        if self._externally_paused:
            self._set_state(ServiceState.PAUSED)
            pdebug("Intern fortgesetzt, aber extern pausiert — bleibe in PAUSED")
        else:
            self._set_state(ServiceState.LISTENING)

    def external_pause(self) -> None:
        """Pausiert Wakeword-Erkennung extern (Server/Plugin)."""
        self._externally_paused = True
        if self._state == ServiceState.LISTENING:
            self._set_state(ServiceState.PAUSED)
        pinfo("Wakeword-Erkennung extern pausiert")

    def external_resume(self) -> None:
        """Setzt extern pausierte Wakeword-Erkennung fort."""
        self._externally_paused = False
        if self._state == ServiceState.PAUSED:
            self._detector.reset()
            self._set_state(ServiceState.LISTENING)
        pinfo("Wakeword-Erkennung extern fortgesetzt")

    async def external_stop(self) -> None:
        """Stoppt Wakeword-Erkennung komplett (Modelle entladen)."""
        self._externally_paused = False
        await self.stop_listening()
        pinfo("Wakeword-Erkennung extern gestoppt")

    async def external_start(self) -> None:
        """Startet Wakeword-Erkennung (Modelle laden + Listening)."""
        if self._state != ServiceState.STOPPED:
            pdebug("Wakeword-Service läuft bereits")
            return
        try:
            self._detector.load_models()
            pinfo(f"Wakeword-Modelle geladen: {self._detector.loaded_models}")
        except Exception as e:
            perror(f"Fehler beim Laden der Wakeword-Modelle: {e}")
            return
        self._detector.on_detection(self._on_wakeword_detected)
        await self.start_listening()
        if self._externally_paused:
            self._set_state(ServiceState.PAUSED)
        pinfo("Wakeword-Erkennung extern gestartet")

    async def _on_manual_trigger(self, event_name: str, event_data: dict) -> None:
        """
        Event-Handler fuer manuellen Wakeword-Trigger (z.B. HID Hook-Taste).

        Ueberspringt die Wakeword-Erkennung und geht direkt in die Aufnahme.
        Arbitration wird forciert (sofortige Auswahl, kein Zeitfenster).
        """
        await self.trigger_manual(
            device=event_data.get("device", "unknown"),
            forced=event_data.get("forced", True),
        )

    async def trigger_manual(self, device: str = "button", forced: bool = True) -> None:
        """
        Manueller Wakeword-Trigger (z.B. durch Tastendruck).

        Ueberspringt die Wakeword-Detection und startet direkt die Aufnahme.
        Mit forced=True wird die Arbitration sofort entschieden (kein Warten
        auf andere Satellites).

        Args:
            device: Name des ausloesenden Geraets
            forced: True = Arbitration sofort entscheiden (kein Fenster)
        """
        if self._state != ServiceState.LISTENING:
            pdebug(f"Manueller Trigger ignoriert — State: {self._state.value}")
            return

        # Cooldown pruefen
        now = time.monotonic()
        if now - self._last_wakeword_time < self._config.wakeword_cooldown_seconds:
            pdebug("Manueller Trigger ignoriert — Cooldown")
            return

        self._last_wakeword_time = now

        pinfo(f"Manueller Wakeword-Trigger ({device}, forced={forced})")

        # Audio-Ducking sofort
        self._duck_audio()

        # Synthetische Detection erstellen (kein echtes Wakeword noetig)
        detection = WakewordDetection(
            wakeword_type=WakewordType.CUSTOM,
            model_name=f"manual:{device}",
            confidence=1.0,      # Maximale Confidence (manuell ausgeloest)
            audio_level=1.0,     # Maximaler Audio-Level (fuer Arbitration-Gewinn)
            timestamp=datetime.now(),
            audio_chunks=[],     # Kein Wakeword-Audio bei manuellem Trigger
        )

        # Pausiere Erkennung
        self._set_state(ServiceState.PAUSED)

        # Session erstellen
        session_id = self._generate_session_id()
        self._current_session = RecordingSession(
            session_id=session_id,
            wakeword_detection=detection,
            start_time=datetime.now(),
        )

        if self._config.standalone_mode:
            # Standalone: Direkt aufnehmen
            self._current_session.is_selected = True
            await self._start_recording()
        else:
            # Client: Arbitration mit Force-Flag
            await self._request_arbitration(detection, forced=forced)

    async def feed_audio(self, audio_data: bytes) -> None:
        """
        Füttert Audio-Daten in den Service (async-API für externen Gebrauch).

        Bei voller Queue wird der älteste Frame verworfen (Ringbuffer-Semantik).

        Args:
            audio_data: 16-bit PCM Audio (16kHz, Mono)
        """
        try:
            self._audio_queue.put_nowait(audio_data)
        except queue.Full:
            try:
                self._audio_queue.get_nowait()
            except queue.Empty:
                pass
            try:
                self._audio_queue.put_nowait(audio_data)
            except queue.Full:
                pass

    # === Audio-Thread (dedizierter Thread, kein asyncio) ===

    def _audio_processing_loop(self) -> None:
        """
        Haupt-Audio-Verarbeitungsschleife — läuft in eigenem Thread.

        Synchron wie in den OpenWakeWord-Beispielen:
        queue.get() blockiert bis Frame da → process_frame() → nächster get().
        Der Thread ist unabhängig vom asyncio-Loop und kann daher nie
        durch Event-Loop-Last verzögert werden.
        """
        while not self._stop_event.is_set():
            try:
                # Blockierend auf nächsten Frame warten (mit Timeout)
                try:
                    audio_data = self._audio_queue.get(timeout=0.1)
                except queue.Empty:
                    continue

                state = self._state

                if state in (ServiceState.LISTENING, ServiceState.PAUSED):
                    # LISTENING: Nur den LETZTEN Frame verarbeiten.
                    # Ältere Frames verwerfen — OpenWakeWord puffert intern
                    # und braucht keine lückenlose Frame-Folge.
                    latest = audio_data
                    skipped = 0
                    while True:
                        try:
                            latest = self._audio_queue.get_nowait()
                            skipped += 1
                        except queue.Empty:
                            break
                    self._process_audio_frame_sync(latest)

                elif state in (ServiceState.RECORDING, ServiceState.WAITING_RESPONSE,
                               ServiceState.FOLLOW_UP):
                    # RECORDING: ALLE Frames verarbeiten (VAD + Audio-Streaming).
                    self._process_audio_frame_sync(audio_data)
                    while True:
                        try:
                            frame = self._audio_queue.get_nowait()
                        except queue.Empty:
                            break
                        self._process_audio_frame_sync(frame)
                        # State könnte sich geändert haben (z.B. → PROCESSING)
                        if self._state not in (ServiceState.RECORDING,
                                               ServiceState.WAITING_RESPONSE,
                                               ServiceState.FOLLOW_UP):
                            break

                # STOPPED/PROCESSING: Frame verwerfen (implizit)

            except Exception as e:
                perror(f"Fehler in Audio-Thread: {e}")

    def _process_audio_frame_sync(self, audio_data: bytes) -> None:
        """
        Verarbeitet einen Audio-Frame synchron im Audio-Thread.

        CPU-intensive Arbeit (Detector, VAD) passiert hier direkt.
        Netzwerk-I/O wird auf den asyncio-Loop dispatcht.
        """
        state = self._state

        if state == ServiceState.LISTENING:
            # Wakeword-Erkennung (synchron — Detector ruft _on_wakeword_detected)
            self._audio_buffer.add_to_pre_buffer(audio_data)
            self._detector.process_frame(audio_data)
            # Server-gesteuerte Aufnahme: streamen via Event-Loop
            if self._server_recording_active and self._audio_streamer:
                self._schedule_async(self._audio_streamer(audio_data))

        elif state == ServiceState.WAITING_RESPONSE:
            # Audio buffern während auf Arbitration gewartet wird
            self._audio_buffer.add_chunk(audio_data)

        elif state == ServiceState.RECORDING:
            # Aufnahme: Buffer + VAD (synchron), Streaming via Event-Loop
            self._audio_buffer.add_chunk(audio_data)
            vad_state = self._vad.process_frame(audio_data)

            if self._audio_streamer:
                self._schedule_async(self._audio_streamer(audio_data))

            # Prüfe ob Aufnahme beendet werden soll.
            # WICHTIG: VADState.SILENCE bedeutet nur "gerade leise", NICHT
            # "Stille-Schwelle erreicht". Daher explizit silence_duration_ms prüfen.
            if vad_state == VADState.TIMEOUT:
                self._set_state(ServiceState.PROCESSING)
                self._schedule_async(self._finish_recording())
            elif vad_state == VADState.NO_SPEECH:
                self._set_state(ServiceState.PROCESSING)
                self._schedule_async(self._finish_recording_no_speech())
            elif (vad_state == VADState.SILENCE
                  and self._vad.has_speech
                  and self._vad.silence_duration_ms >= self._config.silence_timeout_seconds * 1000):
                self._set_state(ServiceState.PROCESSING)
                self._schedule_async(self._finish_recording())

        elif state == ServiceState.FOLLOW_UP:
            # Rückfrage-Aufnahme
            self._audio_buffer.add_chunk(audio_data)
            vad_state = self._vad.process_frame(audio_data)

            if vad_state == VADState.TIMEOUT:
                self._set_state(ServiceState.PROCESSING)
                self._schedule_async(self._finish_follow_up())
            elif (vad_state == VADState.SILENCE
                  and self._vad.has_speech
                  and self._vad.silence_duration_ms >= self._config.silence_timeout_seconds * 1000):
                self._set_state(ServiceState.PROCESSING)
                self._schedule_async(self._finish_follow_up())

        elif state == ServiceState.PAUSED:
            # Pre-Buffer weiter füllen
            self._audio_buffer.add_to_pre_buffer(audio_data)
            # Server-gesteuerte Aufnahme: auch im PAUSED-Zustand streamen
            if self._server_recording_active and self._audio_streamer:
                self._schedule_async(self._audio_streamer(audio_data))

    # === Wakeword-Erkennung und Session-Handling ===

    def _on_wakeword_detected(self, detection: WakewordDetection) -> None:
        """
        Callback wenn Wakeword erkannt wurde.

        Wird aus dem Audio-Thread (via Detector) aufgerufen.

        Args:
            detection: Erkennungs-Details
        """
        if self._state != ServiceState.LISTENING:
            return

        # Cooldown prüfen (verhindert Mehrfach-Erkennung desselben Wakewords)
        now = time.monotonic()
        if now - self._last_wakeword_time < self._config.wakeword_cooldown_seconds:
            pdebug(f"Wakeword ignoriert — Cooldown ({now - self._last_wakeword_time:.1f}s < {self._config.wakeword_cooldown_seconds}s)")
            return

        self._last_wakeword_time = now

        pinfo(
            f"Wakeword erkannt: {detection.model_name} "
            f"(Confidence: {detection.confidence:.2f}, Level: {detection.audio_level:.3f})"
        )

        # Server-gesteuerte Aufnahme auto-stoppen bei Wakeword
        if self._server_recording_active:
            recording_id = self._server_recording_id
            duration = self.stop_server_recording()
            if self._on_server_recording_stopped:
                self._schedule_async(self._on_server_recording_stopped(recording_id, duration))

        # Audio-Ducking: Lautstaerke SOFORT reduzieren (im Audio-Thread)
        self._duck_audio()

        # Pausiere Erkennung (sofort im Audio-Thread)
        self._set_state(ServiceState.PAUSED)

        # Async-Handling auf Event-Loop dispatchen
        self._schedule_async(self._handle_wakeword_detection(detection))

        # User-Callbacks (synchron)
        for callback in self._on_wakeword:
            try:
                callback(detection)
            except Exception:
                pass

    async def _handle_wakeword_detection(self, detection: WakewordDetection) -> None:
        """
        Behandelt Wakeword-Erkennung.

        Läuft auf dem asyncio-Loop (dispatcht vom Audio-Thread).

        Args:
            detection: Erkennungs-Details
        """
        # Generiere Session-ID
        session_id = self._generate_session_id()

        # Erstelle Session
        self._current_session = RecordingSession(
            session_id=session_id,
            wakeword_detection=detection,
            start_time=datetime.now(),
        )

        if self._config.standalone_mode:
            # Standalone: Immer ausgewählt
            self._current_session.is_selected = True
            await self._start_recording()
        else:
            # Client: Arbitration beim Server
            await self._request_arbitration(detection)

    async def _request_arbitration(
        self, detection: WakewordDetection, forced: bool = False,
    ) -> None:
        """
        Sendet Arbitration-Request an Server.

        Args:
            detection: Wakeword-Erkennung
            forced: True = Arbitration sofort entscheiden (kein Fenster warten)
        """
        if not self._send_to_server:
            perror("Keine Server-Verbindung konfiguriert")
            self.resume()
            return

        self._set_state(ServiceState.WAITING_RESPONSE)

        # Starte Aufnahme im Hintergrund (für flüssigen Dialog)
        self._audio_buffer.start_recording(include_pre_buffer=True)

        try:
            # Sende Request
            response = await asyncio.wait_for(
                self._send_to_server("wakeword_detected", {
                    "session_id": self._current_session.session_id,
                    "wakeword_type": detection.wakeword_type.value,
                    "model_name": detection.model_name,
                    "confidence": detection.confidence,
                    "audio_level": detection.audio_level,
                    "audio_chunks": [
                        chunk.hex() for chunk in detection.audio_chunks
                    ],
                    "timestamp": detection.timestamp.isoformat(),
                    "forced": forced,
                }),
                timeout=self._config.arbitration_timeout_seconds,
            )

            if response.get("selected"):
                # Wir wurden ausgewählt
                if not self._current_session:
                    pdebug("Arbitration: Session bereits beendet (Race Condition)")
                    self.resume()
                    return
                self._current_session.is_selected = True
                self._current_session.session_id = response.get(
                    "session_id",
                    self._current_session.session_id,
                )
                await self._start_recording()
            else:
                # Nicht ausgewählt
                pdebug("Arbitration: Nicht ausgewählt")
                self._audio_buffer.clear()
                self._current_session = None
                self.resume()

        except asyncio.TimeoutError:
            pwarn("Arbitration-Timeout")
            self._audio_buffer.clear()
            self._current_session = None
            self.resume()

        except Exception as e:
            perror(f"Arbitration-Fehler: {e}")
            self._audio_buffer.clear()
            self._current_session = None
            self.resume()

    async def _start_recording(self) -> None:
        """Startet die Audio-Aufnahme."""
        self._set_state(ServiceState.RECORDING)

        # VAD starten
        self._vad.start()

        # Buffer starten (falls nicht schon geschehen)
        if not self._audio_buffer.is_recording:
            self._audio_buffer.start_recording(include_pre_buffer=True)

        # Gepuffertes Audio (Pre-Buffer + Wartezeit) an Server senden
        if self._audio_streamer and not self._config.standalone_mode:
            buffered = self._audio_buffer.get_data()
            if buffered:
                await self._audio_streamer(buffered)

        if self._current_session:
            pinfo(f"Aufnahme gestartet (Session: {self._current_session.session_id})")

    async def _finish_recording(self) -> None:
        """Beendet die Aufnahme und sendet Daten."""
        if not self._current_session:
            return

        # State wurde bereits im Audio-Thread auf PROCESSING gesetzt

        # Stoppe Aufnahme
        audio_data = self._audio_buffer.stop_recording()
        self._vad.stop()

        self._current_session.end_time = datetime.now()
        self._current_session.audio_data = audio_data

        pinfo(
            f"Aufnahme beendet: {len(audio_data)} Bytes, "
            f"{self._audio_buffer.duration_seconds:.1f}s"
        )

        # Sende an Server (oder verarbeite lokal)
        await self._send_recording()

    async def _finish_recording_no_speech(self) -> None:
        """Beendet die Aufnahme bei No-Speech (mögliche Fehlauslösung)."""
        if not self._current_session:
            return

        # State wurde bereits im Audio-Thread auf PROCESSING gesetzt

        # Stoppe Aufnahme
        self._audio_buffer.stop_recording()
        self._vad.stop()

        self._current_session.end_time = datetime.now()

        pinfo("Aufnahme abgebrochen — keine Sprache erkannt")

        if not self._config.standalone_mode and self._send_to_server:
            # Client-Modus: RecordingComplete mit speech_detected=False senden
            try:
                await self._send_to_server("recording_complete", {
                    "session_id": self._current_session.session_id,
                    "duration_seconds": self._audio_buffer.duration_seconds,
                    "ended_by": "no_speech",
                    "speech_detected": False,
                    "peak_level": self._vad._peak_level,
                    "vad_stats": self._vad.get_stats(),
                })
            except Exception as e:
                perror(f"Fehler beim Senden: {e}")

        # Session abschließen
        self._complete_session()

    async def _send_recording(self) -> None:
        """Sendet Aufnahme an Server oder verarbeitet lokal."""
        session = self._current_session

        if not session:
            self.resume()
            return

        if self._config.standalone_mode:
            # Lokal: Event emittieren
            await self._application.events.emit("raw_audio_received", {
                "session_id": session.session_id,
                "audio_data": session.audio_data,
                "wakeword_type": session.wakeword_detection.wakeword_type.value,
                "duration_seconds": self._audio_buffer.duration_seconds,
            })

            # Warte auf Verarbeitungs-Ergebnis
            await self._wait_for_processing_result()

            # Standalone: Session sofort abschließen
            self._complete_session()

        else:
            # Client: RecordingComplete Command senden (Audio wurde bereits gestreamt)
            if self._send_to_server:
                # ended_by aus VAD-State ableiten
                vad_state = self._vad.state
                if vad_state == VADState.SILENCE:
                    ended_by = "silence"
                elif vad_state == VADState.TIMEOUT:
                    ended_by = "timeout"
                else:
                    ended_by = "unknown"

                try:
                    await self._send_to_server("recording_complete", {
                        "session_id": session.session_id,
                        "duration_seconds": self._audio_buffer.duration_seconds,
                        "ended_by": ended_by,
                        "speech_detected": self._vad.has_speech,
                        "peak_level": self._vad._peak_level,
                        "vad_stats": self._vad.get_stats(),
                    })
                except Exception as e:
                    perror(f"Fehler beim Senden: {e}")

            # Client-Modus: NICHT sofort zurück zu LISTENING!
            # Bleibe in PROCESSING und warte auf ConversationEnd vom Server.
            # Timeout als Sicherheitsnetz, falls Server nie antwortet.
            pinfo("RecordingComplete gesendet — warte auf ConversationEnd vom Server")
            self._schedule_conversation_end_timeout()

    async def _wait_for_processing_result(self) -> None:
        """Wartet auf Verarbeitungs-Ergebnis (Standalone-Modus)."""
        if not self._current_session:
            return

        session_id = self._current_session.session_id
        result_future: asyncio.Future[dict] = asyncio.Future()

        async def on_processing_result(event_name: str, data) -> None:
            """Handler für processing_result Event."""
            if getattr(data, "session_id", None) == session_id:
                result_future.set_result({
                    "success": getattr(data, "success", True),
                    "text": getattr(data, "text", ""),
                    "intent": getattr(data, "intent", ""),
                    "response_text": getattr(data, "response_text", ""),
                    "error": getattr(data, "error", ""),
                })

        # Event-Manager holen
        event_manager = getattr(self._application, "events", None)
        if not event_manager:
            pwarn("EventManager nicht verfügbar")
            return

        # Temporären Handler registrieren
        event_manager.register("processing_result", on_processing_result)

        try:
            # Auf Ergebnis warten (mit Timeout)
            timeout = self._config.follow_up_timeout_seconds
            result = await asyncio.wait_for(result_future, timeout=timeout)

            if result.get("error"):
                perror(f"Verarbeitungsfehler: {result['error']}")
            else:
                pinfo(f"Verarbeitung abgeschlossen: {result.get('intent', 'unbekannt')}")

        except asyncio.TimeoutError:
            pwarn(f"Timeout beim Warten auf Verarbeitungsergebnis ({timeout}s)")

        finally:
            # Handler wieder entfernen
            event_manager.unregister("processing_result", on_processing_result)

    async def _handle_follow_up(self, response: dict) -> None:
        """
        Behandelt Rückfrage vom Server.

        Args:
            response: Server-Antwort mit Rückfrage
        """
        if not self._current_session:
            return

        self._current_session.follow_up_count += 1
        self._current_session.last_activity = datetime.now()

        pinfo(f"Rückfrage #{self._current_session.follow_up_count}")

        # Starte neue Aufnahme für Antwort
        self._set_state(ServiceState.FOLLOW_UP)
        self._audio_buffer.clear()
        self._audio_buffer.start_recording(include_pre_buffer=True)
        self._vad.start()

    async def _finish_follow_up(self) -> None:
        """Beendet Follow-Up-Aufnahme."""
        if not self._current_session:
            return

        audio_data = self._audio_buffer.stop_recording()
        self._vad.stop()

        # Sende Follow-Up-Antwort
        if self._send_to_server:
            try:
                response = await self._send_to_server("follow_up_response", {
                    "session_id": self._current_session.session_id,
                    "audio_data": audio_data.hex(),
                    "follow_up_number": self._current_session.follow_up_count,
                })

                # Noch eine Rückfrage?
                if response.get("follow_up"):
                    await self._handle_follow_up(response)
                    return

            except Exception as e:
                perror(f"Follow-Up-Fehler: {e}")

        # Session beenden
        self._complete_session()

    def _schedule_conversation_end_timeout(self) -> None:
        """Startet Timeout für ConversationEnd vom Server."""
        async def _timeout_check():
            await asyncio.sleep(15.0)  # 15 Sekunden Timeout
            if self._state == ServiceState.PROCESSING and self._current_session:
                pwarn("ConversationEnd-Timeout — erzwinge Session-Ende")
                self._complete_session()

        self._schedule_async(_timeout_check())

    def _complete_session(self) -> None:
        """Schließt aktuelle Session ab."""
        session = self._current_session

        if session:
            # Callbacks
            for callback in self._on_recording_complete:
                try:
                    callback(session)
                except Exception:
                    pass

        self._current_session = None
        self._audio_buffer.clear()

        # Recording-Config-Override zuruecksetzen (falls temporaer geaendert)
        if hasattr(self, "_original_silence_timeout"):
            self._config.silence_timeout_seconds = self._original_silence_timeout
        if hasattr(self, "_original_max_duration_ms"):
            self._config.vad.max_duration_ms = self._original_max_duration_ms

        # Cooldown erneuern — verhindert Wakeword Re-Detection durch residuales
        # Audio im OpenWakeWord Sliding Window nach dem Resume
        self._last_wakeword_time = time.monotonic()

        # Audio-Queue leeren (verhindert Stau von alten Frames)
        self._drain_audio_queue()

        # Zurück zum Listening
        self.resume()

    async def _on_config_override(self, event_name: str, data) -> None:
        """
        Temporaere Konfigurationsaenderung fuer die naechste Aufnahme.

        Wird von Plugins emittiert um z.B. laengere Aufnahmen (Notizen)
        oder kuerzere Silence-Timeouts zu ermoeglichen.
        Die Werte werden in _complete_session() zurueckgesetzt.
        """
        silence = data.get("silence_timeout_seconds") if isinstance(data, dict) else getattr(data, "silence_timeout_seconds", None)
        max_recording = data.get("max_recording_seconds") if isinstance(data, dict) else getattr(data, "max_recording_seconds", None)

        if silence is not None:
            self._config.silence_timeout_seconds = float(silence)
            pdebug(f"Recording-Config Override: silence_timeout={silence}s")
        if max_recording is not None:
            self._config.vad.max_duration_ms = int(float(max_recording) * 1000)
            pdebug(f"Recording-Config Override: max_recording={max_recording}s")

    def _on_silence_detected(self) -> None:
        """Callback wenn Stille erkannt wurde."""
        pdebug("Stille erkannt")

    def _on_recording_timeout(self) -> None:
        """Callback bei Aufnahme-Timeout."""
        pinfo("Aufnahme-Timeout erreicht")

    def _on_no_speech_detected(self) -> None:
        """Callback wenn keine Sprache erkannt wurde."""
        pinfo("Keine Sprache erkannt — mögliche Fehlauslösung")

    def _generate_session_id(self) -> str:
        """Generiert eine Session-ID."""
        if self._session_id_generator:
            return self._session_id_generator()

        import uuid
        return f"ww-{uuid.uuid4().hex[:12]}"

    # === Public Callbacks ===

    def on_wakeword(self, callback: Callable[[WakewordDetection], None]) -> None:
        """Registriert Callback für Wakeword-Erkennung."""
        self._on_wakeword.append(callback)

    def on_recording_complete(
        self,
        callback: Callable[[RecordingSession], None],
    ) -> None:
        """Registriert Callback für abgeschlossene Aufnahme."""
        self._on_recording_complete.append(callback)

    def on_state_change(self, callback: Callable[[ServiceState], None]) -> None:
        """Registriert Callback für State-Änderungen."""
        self._on_state_change.append(callback)

    def set_server_connection(
        self,
        send_func: Callable[[str, dict], asyncio.Future],
    ) -> None:
        """
        Setzt Server-Kommunikationsfunktion.

        Args:
            send_func: Async-Funktion zum Senden an Server
        """
        self._send_to_server = send_func

    def set_session_id_generator(self, generator: Callable[[], str]) -> None:
        """Setzt Session-ID-Generator."""
        self._session_id_generator = generator

    def set_audio_streamer(self, streamer: Callable[[bytes], Awaitable[bool]]) -> None:
        """
        Setzt Audio-Streaming-Funktion (connection.send_audio).

        Args:
            streamer: Async-Funktion die Audio-Bytes an Server sendet
        """
        self._audio_streamer = streamer

    def setup_for_client(self, connection: "ClientConnection", microphone: "MicrophoneCapture") -> None:
        """
        Konfiguriert Service für Client-Betrieb mit Netzwerk.

        Args:
            connection: ClientConnection zum Server
            microphone: MicrophoneCapture-Instanz
        """
        self._microphone = microphone
        self.set_audio_streamer(connection.send_audio)

        # Future-basierte _send_to_server Funktion
        async def send_to_server(event_type: str, data: dict) -> dict:
            if event_type == "wakeword_detected":
                from trixy_core.network.cmd.wakeword import WakewordDetected
                cmd = WakewordDetected(
                    satellite_id=connection.satellite_id,
                    wakeword_type=data.get("wakeword_type", ""),
                    wakeword_model=data.get("model_name", ""),
                    confidence=data.get("confidence", 0.0),
                    audio_level=data.get("audio_level", 0.0),
                    audio_chunks=data.get("audio_chunks", []),
                    timestamp=data.get("timestamp", ""),
                    client_session_id=data.get("session_id", ""),
                )
                await connection.send_command(cmd)

                # Future für Arbitration-Antwort
                self._arbitration_future = asyncio.get_event_loop().create_future()
                return await self._arbitration_future

            elif event_type == "recording_complete":
                from trixy_core.network.cmd.wakeword import RecordingComplete
                cmd = RecordingComplete(
                    session_id=data.get("session_id", ""),
                    duration_seconds=data.get("duration_seconds", 0.0),
                    ended_by=data.get("ended_by", ""),
                    speech_detected=data.get("speech_detected", True),
                    peak_level=data.get("peak_level", 0.0),
                    vad_stats=data.get("vad_stats", {}),
                )
                await connection.send_command(cmd)
                return {"success": True}

            return {}

        self.set_server_connection(send_to_server)

    async def handle_wakeword_selected(self, data) -> None:
        """
        WakewordSelected vom Server → Future auflösen.

        Args:
            data: WakewordSelected-Daten
        """
        if self._arbitration_future and not self._arbitration_future.done():
            session_id = getattr(data, "session_id", "") or ""
            conversation_id = getattr(data, "conversation_id", "") or ""
            if isinstance(data, dict):
                session_id = data.get("session_id", "")
                conversation_id = data.get("conversation_id", "")
            self._arbitration_future.set_result({
                "selected": True,
                "session_id": session_id or conversation_id,
            })

    async def handle_wakeword_abort(self, data) -> None:
        """
        WakewordAbort vom Server → Future auflösen.

        Args:
            data: WakewordAbort-Daten
        """
        if self._arbitration_future and not self._arbitration_future.done():
            self._arbitration_future.set_result({"selected": False})
        # Cooldown-Timestamp aktualisieren damit nach Abort keine sofortige Re-Detection
        self._last_wakeword_time = time.monotonic()

    async def handle_conversation_end(self) -> None:
        """ConversationEnd → Aufnahme beenden, zurück zu LISTENING."""
        if self._state in (ServiceState.RECORDING, ServiceState.FOLLOW_UP):
            self._vad.stop()
            self._audio_buffer.stop_recording()

        if self._state in (ServiceState.RECORDING, ServiceState.FOLLOW_UP,
                           ServiceState.PROCESSING, ServiceState.WAITING_RESPONSE):
            pinfo("ConversationEnd empfangen — zurück zu LISTENING")
            self._complete_session()

    # === Server-gesteuerte Aufnahme ===

    def start_server_recording(self, recording_id: str) -> None:
        """Aktiviert Audio-Streaming im LISTENING-Zustand."""
        self._server_recording_active = True
        self._server_recording_id = recording_id
        self._server_recording_start_time = time.monotonic()
        pdebug(f"Server-Recording gestartet: {recording_id}")

    def stop_server_recording(self) -> float:
        """Beendet Audio-Streaming. Gibt Dauer zurück."""
        duration = time.monotonic() - self._server_recording_start_time if self._server_recording_active else 0.0
        self._server_recording_active = False
        self._server_recording_id = ""
        self._server_recording_start_time = 0.0
        pdebug(f"Server-Recording gestoppt: {duration:.1f}s")
        return duration

    def set_server_recording_stopped_callback(
        self, cb: Callable[[str, float], Awaitable[None]]
    ) -> None:
        """Setzt Callback für Wakeword-getriggertes Stop."""
        self._on_server_recording_stopped = cb

    def get_stats(self) -> dict:
        """Gibt Statistiken zurück."""
        return {
            "state": self._state.value,
            "detector": self._detector.stats,
            "buffer": self._audio_buffer.get_stats(),
            "vad": self._vad.get_stats() if self._vad.is_active else {},
            "session": {
                "id": self._current_session.session_id if self._current_session else None,
                "follow_up_count": self._current_session.follow_up_count if self._current_session else 0,
            },
        }