Merge pull request #444 from deezer/pad_waveform

Added padding at the begining to avoid tf STFT reconstruction error

spleeter/commands/__init__.py

@@ -170,6 +170,7 @@ def _create_evaluate_parser(parser_factory):
     parser.add_argument('-o', '--output_path', **OPT_OUTPUT)
     parser.add_argument('--mus_dir', **OPT_MUSDB)
     parser.add_argument('-m', '--mwf', **OPT_MWF)
+    parser.add_argument('-B', '--stft-backend', **OPT_STFT_BACKEND)
     return parser
 
 

spleeter/commands/evaluate.py

@@ -77,7 +77,7 @@ def _separate_evaluation_dataset(arguments, musdb_root_directory, params):
             bitrate='128k',
             MWF=arguments.MWF,
             verbose=arguments.verbose,
-            stft_backend="auto"),
+            stft_backend=arguments.stft_backend),
         params)
     return audio_output_directory
 

spleeter/model/__init__.py

@@ -275,9 +275,16 @@ class EstimatorSpecBuilder(object):
         spec_name = self.spectrogram_name
 
         if stft_name not in self._features:
+            # pad input with a frame of zeros
+            waveform =  tf.concat([
+                            tf.zeros((self._frame_length, self._n_channels)),
+                            self._features['waveform']
+                            ],
+                            0
+                        )
             stft_feature = tf.transpose(
                 stft(
-                    tf.transpose(self._features['waveform']),
+                    tf.transpose(waveform),
                     self._frame_length,
                     self._frame_step,
                     window_fn=lambda frame_length, dtype: (
@@ -341,7 +348,7 @@ class EstimatorSpecBuilder(object):
         reshaped = tf.transpose(inversed)
         if time_crop is None:
             time_crop = tf.shape(self._features['waveform'])[0]
-        return reshaped[:time_crop, :]
+        return reshaped[self._frame_length:self._frame_length+time_crop, :]
 
     def _build_mwf_output_waveform(self):
         """ Perform separation with multichannel Wiener Filtering using Norbert.