feat: add spectral rolloff and MFCC delta features to audio extraction

app/services/feature_extractor.py

@@ -189,6 +189,11 @@ def _extract_spectral(y: np.ndarray, sr: int) -> dict:
         y=y, sr=sr, hop_length=_HOP_LENGTH
     )[0]
 
+    # Spectral rolloff — frequency below which 85% of energy
+    rolloff = librosa.feature.spectral_rolloff(
+        y=y, sr=sr, hop_length=_HOP_LENGTH, roll_percent=0.85
+    )[0]
+
     # Spectral contrast — valley-to-peak in frequency bands
     contrast = librosa.feature.spectral_contrast(
         y=y, sr=sr, hop_length=_HOP_LENGTH, n_bands=6
@@ -208,6 +213,10 @@ def _extract_spectral(y: np.ndarray, sr: int) -> dict:
     )
     mel_db = librosa.power_to_db(mel, ref=np.max)
 
+    # MFCC delta (first derivative) and delta-delta (acceleration)
+    mfcc_delta = librosa.feature.delta(mfcc)
+    mfcc_delta2 = librosa.feature.delta(mfcc, order=2)
+
     return {
         "centroid_mean": float(np.mean(centroid)),
         "centroid_std": float(np.std(centroid)),
@@ -215,12 +224,13 @@ def _extract_spectral(y: np.ndarray, sr: int) -> dict:
         "flatness_std": float(np.std(flatness)),
         "bandwidth_mean": float(np.mean(bandwidth)),
         "bandwidth_std": float(np.std(bandwidth)),
+        "rolloff_mean": float(np.mean(rolloff)),
+        "rolloff_std": float(np.std(rolloff)),
         "contrast_mean": float(np.mean(contrast)),
         "contrast_std": float(np.std(contrast)),
         "mfcc_variance": float(np.mean(np.var(mfcc, axis=1))),
-        "mfcc_delta_var": float(np.mean(np.var(
-            librosa.feature.delta(mfcc), axis=1
-        ))),
+        "mfcc_delta_var": float(np.mean(np.var(mfcc_delta, axis=1))),
+        "mfcc_delta2_var": float(np.mean(np.var(mfcc_delta2, axis=1))),
         "rms_mean": float(np.mean(rms)),
         "rms_std": float(np.std(rms)),
         "mel_flatness": float(np.mean(np.std(mel_db, axis=0))),