Update

app.py

@@ -6,20 +6,15 @@ import traceback
 
 def detect_watermark(audio_file_path):
     try:
-        # Load the audio file
+        # Load the audio file and resample if necessary
         waveform, sample_rate = torchaudio.load(audio_file_path)
+        if sample_rate != 16000:
+            resample_transform = torchaudio.transforms.Resample(orig_freq=sample_rate, new_freq=16000)
+            waveform = resample_transform(waveform)
+            sample_rate = 16000
 
-        # Normalize the waveform to be in the range [-1, 1]
-        waveform_max = torch.max(torch.abs(waveform))
-        if waveform_max > 0:
-            waveform = waveform / waveform_max
-
-        # Resample the waveform to 16kHz if necessary
-        target_sample_rate = 16000
-        if sample_rate != target_sample_rate:
-            resampler = torchaudio.transforms.Resample(orig_freq=sample_rate, new_freq=target_sample_rate)
-            waveform = resampler(waveform)
-            sample_rate = target_sample_rate
+        # Normalize waveform loudness
+        waveform = torch.clamp(waveform, min=-1.0, max=1.0)
 
         # Ensure waveform has a batch dimension for processing
         if waveform.ndim < 3:
@@ -28,18 +23,16 @@ def detect_watermark(audio_file_path):
         # Initialize the AudioSeal detector
         detector = AudioSeal.load_detector("audioseal_detector_16bits")
         
-        # Set a conservative threshold for watermark detection
-        message_threshold = 0.99  # A higher threshold means more confidence is required to classify as AI-generated
-        
-        result, confidence = detector.detect_watermark(waveform, message_threshold=message_threshold)
+        # Detect watermark (simplified to binary outcome for AI-generated or not)
+        result, _ = detector.detect_watermark(waveform, message_threshold=0.99)
 
-        # Interpret the detection result
-        if result:
-            detection_result = f"AI-generated with confidence {confidence}"
-        else:
-            detection_result = "Genuine or the AI watermark is undetectable at the current threshold"
+        # Simplify the output message
+        if result == 1:  # Assuming '1' means AI-generated
+            detection_result = "The audio is likely AI-generated."
+        else:  # Assuming '0' means human-created
+            detection_result = "The audio is likely human-created."
 
-        return f"This audio is likely {detection_result}."
+        return detection_result
     except Exception as e:
         error_traceback = traceback.format_exc()
         return f"Error occurred: {e}\n\n{error_traceback}"