Fix scaling

inference.py

@@ -47,7 +47,6 @@ class EnClap:
             "pad_token_id": 1,
             "max_length": 50,
         }
-        self.scale_factor = 2**15
         self.max_seq_len = config.max_position_embeddings - 3
 
     @torch.no_grad()
@@ -65,8 +64,10 @@ class EnClap:
     ) -> str:
         if generation_config is None:
             generation_config = self.generation_config
-        if audio.dtype == torch.int or audio.dtype == torch.short:
-            audio = audio / self.scale_factor
+        if audio.dtype == torch.short:
+            audio = audio / 2**15
+        if audio.dtype == torch.int:
+            audio = audio / 2**31
         encodec_audio = (
             convert_audio(
                 audio.unsqueeze(0), res, self.encodec.sample_rate, self.encodec.channels