Update custom_interface.py

custom_interface.py

@@ -85,14 +85,64 @@ class ASR(Pretrained):
         return seq
     
 
-    def classify_file(self, path):
-        # waveform = self.load_audio(path)
-        waveform, sr = librosa.load(path, sr=16000)
-        waveform = torch.tensor(waveform)
+    # def classify_file(self, path):
+    #     # waveform = self.load_audio(path)
+    #     waveform, sr = librosa.load(path, sr=16000)
+    #     waveform = torch.tensor(waveform)
 
-        # Fake a batch:
-        batch = waveform.unsqueeze(0)
-        rel_length = torch.tensor([1.0])
-        outputs = self.encode_batch(batch, rel_length)
+    #     # Fake a batch:
+    #     batch = waveform.unsqueeze(0)
+    #     rel_length = torch.tensor([1.0])
+    #     outputs = self.encode_batch(batch, rel_length)
        
+    #     return outputs
+
+    def classify_file(self, path):
+        # Load the audio file
+        waveform, sr = librosa.load(path, sr=16000)
+    
+        # Get audio length in seconds
+        audio_length = len(waveform) / sr
+        print(f"Audio length: {audio_length:.2f} seconds")
+    
+        # Detect non-silent segments
+        non_silent_intervals = librosa.effects.split(waveform, top_db=20)  # Adjust top_db for sensitivity
+    
+        segments = []
+        current_segment = []
+        current_length = 0
+        max_duration = 20 * sr  # Maximum segment duration in samples (20 seconds)
+    
+        for interval in non_silent_intervals:
+            start, end = interval
+            segment_part = waveform[start:end]
+    
+            # If adding the next part exceeds max duration, store the segment and start a new one
+            if current_length + len(segment_part) > max_duration:
+                segments.append(np.concatenate(current_segment))
+                current_segment = []
+                current_length = 0
+    
+            current_segment.append(segment_part)
+            current_length += len(segment_part)
+    
+        # Append the last segment if it's not empty
+        if current_segment:
+            segments.append(np.concatenate(current_segment))
+    
+        # Process each segment
+        outputs = []
+        for i, segment in enumerate(segments):
+            print(f"Processing segment {i + 1}/{len(segments)}, length: {len(segment) / sr:.2f} seconds")
+    
+            segment_tensor = torch.tensor(segment)
+    
+            # Fake a batch for the segment
+            batch = segment_tensor.unsqueeze(0)
+            rel_length = torch.tensor([1.0])  # Adjust if necessary
+    
+            # Pass the segment through the ASR model
+            segment_output = self.encode_batch(batch, rel_length)
+            outputs.append(segment_output)
+    
         return outputs