Update audio_spec.py

Fix the potential "nan" problem of the preprocess_model used by BrowserFftSpec.

The preprocess_model will produce "nan" results when input tensors contain too many zero entries. Such inputs are common because there may be delays in recording devices resulting in an empty stream at the beginning of the audio file. This problem can be fixed by adding small Gaussian noise to the inputs before they go through the preprocess_model.

Author: Yijia Shao

tensorflow_examples/lite/model_maker/core/task/model_spec/audio_spec.py

@@ -283,7 +283,8 @@ def _split(self, wav, label):
   ])
   def _preprocess(self, x, label):
     """Preprocess the dataset to extract the spectrum."""
-    # Add small Gaussian noise to the input x to solve the potential "nan" problem of the preprocess_model
+    # Add small Gaussian noise to the input x
+    # to solve the potential "nan" problem of the preprocess_model
     x = x + 1e-05 * tf.random.normal(x.shape)
     # x has shape (1, EXPECTED_WAVEFORM_LENGTH)
     spectrum = self._preprocess_model(x)