Update custom_interface_app.py

custom_interface_app.py

@@ -231,6 +231,48 @@ class ASR(Pretrained):
             return outputs
 
 
+    def classify_file_whisper_mkd_streaming(self, waveform, device):
+        # Load the audio file
+        # waveform, sr = librosa.load(path, sr=16000)
+
+        # Get audio length in seconds
+        audio_length = len(waveform) / 16000
+        
+        if audio_length >= 20:
+            # split audio every 20 seconds
+            segments = []
+            max_duration = 20 * 16000  # Maximum segment duration in samples (20 seconds)
+            num_segments = int(np.ceil(len(waveform) / max_duration))
+            start = 0
+            for i in range(num_segments):
+                end = start + max_duration
+                if end > len(waveform):
+                    end = len(waveform)
+                segment_part = waveform[start:end]
+                segment_len = len(segment_part) / 16000
+                if segment_len < 1:
+                    continue
+                segments.append(segment_part)
+                start = end
+
+            for segment in segments:
+                segment_tensor = torch.tensor(segment).to(device)
+
+                # Fake a batch for the segment
+                batch = segment_tensor.unsqueeze(0).to(device)
+                rel_length = torch.tensor([1.0]).to(device)
+
+                # Pass the segment through the ASR model
+                segment_output = self.encode_batch_whisper(device, batch, rel_length)
+                yield segment_output
+        else:
+            waveform = torch.tensor(waveform).to(device)
+            waveform = waveform.to(device)
+            batch = waveform.unsqueeze(0)
+            rel_length = torch.tensor([1.0]).to(device)
+            outputs = self.encode_batch_whisper(device, batch, rel_length)
+            yield outputs
+
 
     def classify_file_whisper(self, waveform, pipe, device):
         # waveform, sr = librosa.load(path, sr=16000)