# backend/voice/wake_word.py
# Real wake word detection using openWakeWord
# In CLOUD_ENV mode this is a no-op stub — HF Spaces has no microphone.
# On local PC/device, it auto-downloads tflite models and listens for both personas.
import os
import logging
import numpy as np
from typing import Optional

IS_CLOUD = os.environ.get("CLOUD_ENV", "false").lower() == "true"

class WakeWordDetector:
    """
    Dual-persona wake word detector.
    - Local mode  : listens for 'hey jarvis' AND 'hey friday', auto-downloads models.
    - Cloud mode  : no-op stub (HF Spaces has no microphone — mobile/PC handles wake word).
    """
    def __init__(self, keywords: list[str] = None):
        self.mute_wake_word_during_tts = False
        self._enabled = False
        self.model = None

        if IS_CLOUD:
            logging.info("[WakeWord] Cloud mode — wake word detection disabled (no mic on server).")
            return

        if keywords is None:
            keywords = ["hey jarvis", "hey friday"]

        try:
            import openwakeword
            from openwakeword.model import Model
            # Download all built-in pre-trained models on first run
            openwakeword.utils.download_models()
            # Load only the keywords that have matching tflite files
            import openwakeword.utils as oww_utils
            available = oww_utils.get_pretrained_model_paths()
            available_names = {os.path.splitext(os.path.basename(p))[0].replace("_", " ").lower(): p for p in available}
            loadable = []
            for kw in keywords:
                clean = kw.lower().replace(" ", "_")
                matched = [p for name, p in available_names.items() if clean in name or kw.lower() in name]
                if matched:
                    loadable.append(matched[0])
                else:
                    logging.warning(f"[WakeWord] No pretrained model found for '{kw}' — skipping.")
            if loadable:
                self.model = Model(wakeword_models=loadable, inference_framework="tflite")
                self._enabled = True
                logging.info(f"[WakeWord] Loaded {len(loadable)} wake word model(s): {loadable}")
            else:
                logging.warning("[WakeWord] No wake word models loaded — all keywords unsupported.")
        except Exception as e:
            logging.warning(f"[WakeWord] Failed to load wake word models: {e}. Continuing without wake word.")

    def process_frame(self, audio_frame: np.ndarray) -> Optional[str]:
        if not self._enabled or self.model is None:
            return None
        if getattr(self, 'mute_wake_word_during_tts', False):
            return None
        try:
            predictions = self.model.predict(audio_frame)
            for keyword, score in predictions.items():
                if score > 0.5:
                    return keyword
        except Exception as e:
            import logging; logging.getLogger(__name__).error(f"Swallowed exception: {e}")
        return None