synthetic_detector.py

import numpy as np
import librosa


def detect_synthetic_voice(y, sr, spectral):
    """
    Compute a heuristic synthetic probability score.
    This is NOT a classifier — just an informational hint.

    Uses:
    - MFCC variance (synthetic voices are extremely stable)
    - HF spectral symmetry (AI vocoders often show smooth HF rolloff)
    - Pitch jitter (AI voices have ~0 jitter)
    """

    # ============================================================
    # 1. MFCC Stability
    # ============================================================

    try:
        mfcc = librosa.feature.mfcc(y=y, sr=sr, n_mfcc=13)
        mfcc_var = float(np.mean(np.var(mfcc, axis=1)))
    except Exception:
        mfcc_var = 0.0

    # Human speech → high variance
    # Synthetic → overly stable
    mfcc_score = 1.0 - np.tanh(mfcc_var / 50.0)

    # ============================================================
    # 2. Pitch Jitter
    # ============================================================

    try:
        f0 = librosa.yin(y, fmin=80, fmax=400)
        f0 = f0[np.isfinite(f0)]

        if len(f0) > 5:
            jitter = float(np.std(f0) / np.mean(f0))
        else:
            jitter = 0.0
    except Exception:
        jitter = 0.0

    # Human → jitter = 1–5%  
    # AI → jitter close to 0  
    jitter_score = 1.0 - np.tanh(jitter * 10)

    # ============================================================
    # 3. HF Smoothness / Symmetry
    # ============================================================

    hf_env = spectral.get("hf_env", None)
    if hf_env is not None:
        # Measure average change between bins
        diffs = np.abs(np.diff(hf_env))
        smoothness = 1.0 - np.tanh(np.mean(diffs) / 5.0)
    else:
        smoothness = 0.0

    # ============================================================
    # Combine Scores
    # ============================================================

    prob = float(np.clip((mfcc_score + jitter_score + smoothness) / 3.0, 0, 1))

    label = "AI" if prob >= 0.55 else "Human"

    return {
        "synthetic_probability": prob,
        "synthetic_label": label,
        "mfcc_variance": mfcc_var,
        "pitch_jitter": jitter,
        "hf_smoothness": float(smoothness)
    }