Update pipeline.py

pipeline.py

@@ -229,38 +229,60 @@ audio_label_map = {
     1: "Fake audio"
 }
 
-NB_SAMP = 64600  # Expected sample length by RawNet2
+RAWNET_SAMPLE_RATE = 16000  # RawNet2 was trained on 16kHz audio — never change this
+NB_SAMP = 64600             # Exactly 4.0375 seconds at 16kHz
 
 def deepfakes_audio_predict(input_audio):
     """
-    Gradio gr.Audio() returns a tuple: (sample_rate, numpy_array)
-    numpy_array is int16 by default and needs float32 normalization.
+    Gradio gr.Audio() returns a tuple: (sample_rate, numpy_array).
+    
+    Critical fixes applied:
+      1. Resample to RAWNET_SAMPLE_RATE (16000 Hz) — the model was trained at 16kHz.
+         Without this, a 44100Hz input has its first ~1.46s fed to a model expecting ~4s,
+         completely breaking the SincConv filterbank frequency assumptions.
+      2. Stereo → mono before resampling (librosa.resample requires 1D input).
+      3. Normalize AFTER resampling to avoid float64 precision issues from librosa.
     """
     sr, x = input_audio
 
+    print(f"[Audio] Input sample rate: {sr} Hz, samples: {len(x)}, dtype: {x.dtype}")
+
+    # Step 1: Convert to float32
     x = x.astype(np.float32)
-    if x.max() > 1.0:
-        x = x / 32768.0  # Normalize int16 range to float32
 
-    # If stereo, convert to mono by averaging channels
+    # Step 2: Normalize int16 → [-1.0, 1.0] range
+    if np.abs(x).max() > 1.0:
+        x = x / 32768.0
+
+    # Step 3: Stereo → mono (must be done before librosa.resample)
     if x.ndim == 2:
         x = x.mean(axis=1)
 
-    # RawNet2 expects exactly nb_samp=64600 samples — pad or trim
+    # Step 4: Resample to 16000 Hz — THIS WAS THE ROOT CAUSE BUG
+    # RawNet2's SincConv filterbank is hard-coded to 16kHz frequencies.
+    # Feeding audio at any other sample rate produces completely wrong filter responses.
+    if sr != RAWNET_SAMPLE_RATE:
+        print(f"[Audio] Resampling from {sr} Hz → {RAWNET_SAMPLE_RATE} Hz")
+        x = librosa.resample(x, orig_sr=sr, target_sr=RAWNET_SAMPLE_RATE)
+        print(f"[Audio] After resample: {len(x)} samples ({len(x)/RAWNET_SAMPLE_RATE:.2f}s)")
+
+    # Step 5: Pad or trim to exactly NB_SAMP (64600) samples
     if len(x) < NB_SAMP:
         x = np.pad(x, (0, NB_SAMP - len(x)), mode='constant')
     else:
         x = x[:NB_SAMP]
 
-    # Convert to tensor with batch dimension: [1, nb_samp]
+    # Step 6: Build tensor [1, NB_SAMP] and run inference
     x_pt = torch.tensor(x, dtype=torch.float32).unsqueeze(0)
 
     audio_model = load_audio_model()
 
     with torch.no_grad():
-        grads = audio_model(x_pt)
+        logits = audio_model(x_pt)
+
+    logits_np = logits.detach().numpy()
+    result = np.argmax(logits_np)
 
-    grads_np = grads.detach().numpy()
-    result = np.argmax(grads_np)
+    print(f"[Audio] Logits: {logits_np}, Predicted class: {result} ({audio_label_map[int(result)]})")
 
     return audio_label_map[int(result)]