Upload app.py with huggingface_hub

app.py

@@ -179,15 +179,25 @@ def extract3ch(audio, rate=SR):
     mel = librosa.feature.melspectrogram(
         y=audio, sr=rate, n_mels=NMELS, n_fft=NFFT, hop_length=HOP)
     meldb = librosa.power_to_db(mel, ref=np.max)
+    # Compute deltas on the ORIGINAL time resolution (before zoom)
+    d1 = librosa.feature.delta(meldb, order=1)
+    d2 = librosa.feature.delta(meldb, order=2)
+    # Now zoom all 3 channels to MAXFRAMES
     if meldb.shape[1] != MAXFRAMES:
-        meldb = scipyzoom(meldb, (1, MAXFRAMES / meldb.shape[1]), order=1)
+        zf = MAXFRAMES / meldb.shape[1]
+        meldb = scipyzoom(meldb, (1, zf), order=1)
+        d1 = scipyzoom(d1, (1, zf), order=1)
+        d2 = scipyzoom(d2, (1, zf), order=1)
     if meldb.shape[1] > MAXFRAMES:
         meldb = meldb[:, :MAXFRAMES]
+        d1 = d1[:, :MAXFRAMES]
+        d2 = d2[:, :MAXFRAMES]
     elif meldb.shape[1] < MAXFRAMES:
-        meldb = np.pad(meldb, ((0, 0), (0, MAXFRAMES - meldb.shape[1])),
+        pad_w = MAXFRAMES - meldb.shape[1]
+        meldb = np.pad(meldb, ((0, 0), (0, pad_w)),
                        constant_values=meldb.min())
-    d1 = librosa.feature.delta(meldb, order=1)
-    d2 = librosa.feature.delta(meldb, order=2)
+        d1 = np.pad(d1, ((0, 0), (0, pad_w)), constant_values=0)
+        d2 = np.pad(d2, ((0, 0), (0, pad_w)), constant_values=0)
     return np.stack([meldb, d1, d2], axis=0).astype(np.float32)
 
 # ── Prediction Function ─────────────────────────────────────────────