scripts/02_acoustic_rf.py

from pathlib import Path
import numpy as np
import pandas as pd
import librosa
from sklearn.ensemble import ExtraTreesClassifier
from sklearn.preprocessing import StandardScaler
from sklearn.model_selection import train_test_split
from sklearn.metrics import roc_auc_score, confusion_matrix, ConfusionMatrixDisplay
import matplotlib
matplotlib.use("Agg")
import matplotlib.pyplot as plt
import warnings
warnings.filterwarnings("ignore")

DATA_DIR = Path("output/voiceguard")
AUDIO_DIR = DATA_DIR
SR = 16_000
DURATION = 4
N_SAMPLES = SR * DURATION


train_df = pd.read_csv(DATA_DIR / "train.csv")
train_df["label"] = (train_df["label"] == "fake").astype(int)
test_df  = pd.read_csv(DATA_DIR / "test.csv")
print("Train:", train_df.shape, "  Test:", test_df.shape)


def load_audio(path, sr=SR, n_samples=N_SAMPLES):
    y, _ = librosa.load(path, sr=sr, mono=True)
    if len(y) < n_samples:
        y = np.pad(y, (0, n_samples - len(y)))
    return y[:n_samples]


def hnr_autocorr(y, sr=SR, fmin=75, fmax=400):
    frame_len = int(sr * 0.04)
    hop = int(sr * 0.01)
    hnrs = []
    for start in range(0, len(y) - frame_len, hop):
        frame = y[start: start + frame_len] * np.hanning(frame_len)
        r = np.correlate(frame, frame, mode="full")
        r = r[len(r) // 2:]
        r0 = r[0] + 1e-8
        min_lag, max_lag = int(sr / fmax), int(sr / fmin)
        if max_lag >= len(r):
            continue
        r_max = r[min_lag:max_lag].max()
        hnrs.append(10 * np.log10(r_max / (r0 - r_max + 1e-8)))
    return float(np.mean(hnrs)) if hnrs else 0.0


def jitter_shimmer_approx(y, sr=SR):
    """Approximate jitter (F0 variation) and shimmer (amplitude variation)."""
    # F0 via yin (deterministic, much faster than pyin)
    try:
        f0 = librosa.yin(y, fmin=75, fmax=400, sr=sr,
                         frame_length=int(sr * 0.04),
                         hop_length=int(sr * 0.01))
        f0_voiced = f0[(f0 > 75) & (f0 < 400)]
    except Exception:
        f0_voiced = np.array([])

    jitter = float(np.std(np.diff(f0_voiced))) if len(f0_voiced) > 1 else 0.0

    # amplitude on voiced frames
    hop = int(sr * 0.01)
    frame_len = int(sr * 0.04)
    amps = []
    for start in range(0, len(y) - frame_len, hop):
        amps.append(float(np.sqrt(np.mean(y[start: start + frame_len] ** 2))))
    amps = np.array(amps)
    if len(amps) > 1 and amps.mean() > 1e-8:
        shimmer = float(np.std(np.diff(amps)) / amps.mean())
    else:
        shimmer = 0.0
    return jitter, shimmer


def extract_features(path):
    y = load_audio(path)
    n_fft = 512

    sf   = librosa.feature.spectral_flatness(y=y)[0]
    mfcc = librosa.feature.mfcc(y=y, sr=SR, n_mfcc=20)
    mfcc_delta = librosa.feature.delta(mfcc)
    zcr  = librosa.feature.zero_crossing_rate(y)[0]
    rms  = librosa.feature.rms(y=y)[0]

    chroma = librosa.feature.chroma_stft(y=y, sr=SR, n_fft=n_fft)

    cent  = librosa.feature.spectral_centroid(y=y, sr=SR, n_fft=n_fft)[0]
    bw    = librosa.feature.spectral_bandwidth(y=y, sr=SR, n_fft=n_fft)[0]
    roll  = librosa.feature.spectral_rolloff(y=y, sr=SR, n_fft=n_fft)[0]

    S = np.abs(librosa.stft(y, n_fft=n_fft)) ** 2
    freqs = librosa.fft_frequencies(sr=SR, n_fft=n_fft)
    hfe = float(S[freqs >= 4000].sum() / (S.sum() + 1e-8))

    jitter, shimmer = jitter_shimmer_approx(y)
    hnr = hnr_autocorr(y)

    feat = np.concatenate([
        [np.mean(sf), np.std(sf)],
        [hfe],
        np.mean(mfcc, axis=1),
        np.std(mfcc, axis=1),
        np.mean(mfcc_delta, axis=1),
        [np.mean(zcr), np.std(zcr)],
        [np.mean(rms), np.std(rms)],
        np.mean(chroma, axis=1),
        [np.mean(cent), np.std(cent)],
        [np.mean(bw),   np.std(bw)],
        [np.mean(roll), np.std(roll)],
        [jitter, shimmer, hnr],
    ])
    return feat


from joblib import Parallel, delayed
from tqdm import tqdm
Path("cache").mkdir(exist_ok=True)

def _ex(row_id):
    return extract_features(AUDIO_DIR / row_id)

def cached_extract(df, cache_name):
    cp = Path("cache") / cache_name
    if cp.exists():
        print(f"[cache] Loading {cache_name}"); return np.load(cp)
    feats = np.array(Parallel(n_jobs=-1, prefer="threads")(
        delayed(_ex)(r["id"]) for _, r in tqdm(df.iterrows(), total=len(df))))
    np.save(cp, feats); print(f"[cache] Saved {cache_name}"); return feats

print("Extracting train features …")
X_train_raw = np.nan_to_num(cached_extract(train_df, "vg_02_train.npy"), nan=0.0, posinf=0.0, neginf=0.0)
y_train = train_df["label"].to_numpy()
print("Done. Shape:", X_train_raw.shape)

print("Extracting test features …")
X_test_raw = np.nan_to_num(cached_extract(test_df, "vg_02_test.npy"), nan=0.0, posinf=0.0, neginf=0.0)
print("Done. Shape:", X_test_raw.shape)


X_tr, X_val, y_tr, y_val = train_test_split(
    X_train_raw, y_train, test_size=0.2, random_state=42, stratify=y_train
)
scaler = StandardScaler()
X_tr_s   = scaler.fit_transform(X_tr)
X_val_s  = scaler.transform(X_val)
X_test_s = scaler.transform(X_test_raw)


clf = ExtraTreesClassifier(n_estimators=300, random_state=42, n_jobs=-1)
clf.fit(X_tr_s, y_tr)

val_probs = clf.predict_proba(X_val_s)[:, 1]
val_preds = (val_probs >= 0.5).astype(int)
auroc = roc_auc_score(y_val, val_probs)
print(f"Validation AUROC: {auroc:.4f}")


importances = clf.feature_importances_
top_idx = np.argsort(importances)[::-1][:20]
fig, ax = plt.subplots(figsize=(9, 4))
ax.bar(range(20), importances[top_idx])
ax.set_xticks(range(20))
ax.set_xticklabels([f"f{i}" for i in top_idx], rotation=45)
ax.set_ylabel("Importance")
ax.set_title("Top-20 Feature Importances")
plt.tight_layout()
plt.savefig("/dev/null")


cm = confusion_matrix(y_val, val_preds)
disp = ConfusionMatrixDisplay(cm, display_labels=["real", "fake"])
fig, ax = plt.subplots(figsize=(4, 4))
disp.plot(ax=ax, colorbar=False)
ax.set_title(f"Confusion Matrix  AUROC={auroc:.3f}")
plt.tight_layout()
plt.savefig("/dev/null")


test_probs = clf.predict_proba(X_test_s)[:, 1]
np.save("probs_rf.npy", test_probs)

submission = pd.DataFrame({"id": test_df["id"], "score": test_probs})
submission.to_csv("submission_rf.csv", index=False)
print(submission.head())
print("Saved submission_rf.csv  |  probs_rf.npy")