removed redundant torch tensor copy

pipeline.py

@@ -53,9 +53,10 @@ class PreTrainedPipeline():
             sampling_rate=self.sampling_rate
         )['input_values']
 
+        input_values = input_values.to(self.device)
+
         with torch.no_grad():
-            # input_values should be a 1D numpy array by now
-            input_values = torch.tensor(input_values, device=self.device)
+            # input_values should be a 2D tensor by now. 1st dim represents audio channels.
             model_outs = self.model(input_values)
         logits = model_outs.logits.cpu().detach().numpy()