diff --git a/nn_fingerprint.py b/nn_fingerprint.py
index ac07b0b..27925f3 100644
--- a/nn_fingerprint.py
+++ b/nn_fingerprint.py
@@ -64,8 +64,8 @@ class NNFingerprint(NNOperator):
         if model_path.endswith('.onnx'):
             log.warning('Using onnx.')
             self.model = onnxruntime.InferenceSession(
-            model_path,
-            providers=['CPUExecutionProvider'] if self.device == 'cpu' else ['CUDAExecutionProvider']
+                model_path,
+                providers=['CPUExecutionProvider'] if self.device == 'cpu' else ['CUDAExecutionProvider']
             )
         else:
             try:
@@ -120,7 +120,7 @@ class NNFingerprint(NNOperator):
                 audio = numpy.hstack(frames)
                 if len(audio.shape) == 1:
                     audio = audio[None, :]
-            audio = self.int2float(audio)
+            audio = self.int2float(audio, dtype='float32')
             audio = torch.from_numpy(audio)
         assert len(audio.shape) == 2
 
@@ -158,10 +158,15 @@ class NNFingerprint(NNOperator):
         assert dtype.kind == 'f'
 
         if wav.dtype.kind in 'iu':
-            ii = numpy.iinfo(wav.dtype)
-            abs_max = 2 ** (ii.bits - 1)
-            offset = ii.min + abs_max
-            return (wav.astype(dtype) - offset) / abs_max
+            # ii = numpy.iinfo(wav.dtype)
+            # abs_max = 2 ** (ii.bits - 1)
+            # offset = ii.min + abs_max
+            # return (wav.astype(dtype) - offset) / abs_max
+            if wav.dtype != 'int16':
+                wav = (wav >> 16).astype(numpy.int16)
+            assert wav.dtype == 'int16'
+            wav = (wav / 32768.0).astype(dtype)
+            return wav
         else:
             log.warning('Converting float dtype from %s to %s.', wav.dtype, dtype)
             return wav.astype(dtype)