# app/model.py
import joblib
import numpy as np
import librosa
import os

# Global placeholders
model = None
scaler = None

# ==========================
# Load Model and Scaler
# ==========================
def load_models(model_path="emotion_model.pkl", scaler_path="scaler.pkl"):
    global model, scaler
    if not os.path.exists(model_path) or not os.path.exists(scaler_path):
        raise FileNotFoundError("Model or scaler .pkl files not found. Upload them to the app directory.")
    model = joblib.load(model_path)
    scaler = joblib.load(scaler_path)
    print("✅ Tone-based emotion model loaded successfully.")

# ==========================
# Feature extraction
# ==========================
def extract_features(audio_path):
    """
    Extract tone-based features from audio:
    - MFCC
    - Pitch
    - Energy
    """
    audio, sr = librosa.load(audio_path, sr=16000)
    mfcc = librosa.feature.mfcc(y=audio, sr=sr, n_mfcc=13)
    mfcc_mean = np.mean(mfcc, axis=1)
    pitches, magnitudes = librosa.piptrack(y=audio, sr=sr)
    pitch = np.mean(pitches[pitches > 0]) if np.any(pitches > 0) else 0
    energy = np.mean(librosa.feature.rms(y=audio))
    return np.hstack([mfcc_mean, pitch, energy])

# ==========================
# Predict Emotion
# ==========================
def predict_tone(audio_path):
    global model, scaler
    if model is None or scaler is None:
        raise RuntimeError("Model and scaler must be loaded first.")

    features = extract_features(audio_path).reshape(1, -1)
    features_scaled = scaler.transform(features)
    pred_label = model.predict(features_scaled)[0]
    pred_proba = model.predict_proba(features_scaled)[0]

    # Confidence of the predicted class
    pred_index = np.argmax(pred_proba)
    confidence = float(pred_proba[pred_index])

    return {
        "emotion_label": pred_label,
        "confidence": confidence,
        "note": "Tone-based prediction (less text bias)"
    }