Merge pull request #50 from keithito/trim-audio

Trim audio at the first silence

demo_server.py

@@ -86,7 +86,6 @@ if __name__ == '__main__':
     help='Hyperparameter overrides as a comma-separated list of name=value pairs')
   args = parser.parse_args()
   os.environ['TF_CPP_MIN_LOG_LEVEL'] = '2'
-  hparams.max_iters = 100
   hparams.parse(args.hparams)
   print(hparams_debug_string())
   synthesizer.load(args.checkpoint)

eval.py

@@ -45,7 +45,6 @@ def main():
     help='Hyperparameter overrides as a comma-separated list of name=value pairs')
   args = parser.parse_args()
   os.environ['TF_CPP_MIN_LOG_LEVEL'] = '2'
-  hparams.max_iters = 100
   hparams.parse(args.hparams)
   run_eval(args)
 

synthesizer.py

@@ -2,6 +2,7 @@ import io
 import numpy as np
 import tensorflow as tf
 from hparams import hparams
+from librosa import effects
 from models import create_model
 from text import text_to_sequence
 from util import audio
@@ -32,6 +33,8 @@ class Synthesizer:
       self.model.input_lengths: np.asarray([len(seq)], dtype=np.int32)
     }
     wav = self.session.run(self.wav_output, feed_dict=feed_dict)
+    wav = audio.inv_preemphasis(wav)
+    wav = wav[:audio.find_endpoint(wav)]
     out = io.BytesIO()
-    audio.save_wav(audio.inv_preemphasis(wav), out)
+    audio.save_wav(wav, out)
     return out.getvalue()

util/audio.py

@@ -52,6 +52,16 @@ def melspectrogram(y):
   return _normalize(S)
 
 
+def find_endpoint(wav, threshold_db=-40, min_silence_sec=0.8):
+  window_length = int(hparams.sample_rate * min_silence_sec)
+  hop_length = int(window_length / 4)
+  threshold = _db_to_amp(threshold_db)
+  for x in range(hop_length, len(wav) - window_length, hop_length):
+    if np.max(wav[x:x+window_length]) < threshold:
+      return x + hop_length
+  return len(wav)
+
+
 def _griffin_lim(S):
   '''librosa implementation of Griffin-Lim
   Based on https://github.com/librosa/librosa/issues/434
@@ -111,7 +121,6 @@ def _stft_parameters():
 # Conversions:
 
 _mel_basis = None
-_inv_mel_basis = None
 
 def _linear_to_mel(spectrogram):
   global _mel_basis
@@ -119,12 +128,6 @@ def _linear_to_mel(spectrogram):
     _mel_basis = _build_mel_basis()
   return np.dot(_mel_basis, spectrogram)
 
-def _mel_to_linear(mel_spectrogram):
-  global _inv_mel_basis
-  if _inv_mel_basis is None:
-    _inv_mel_basis = np.linalg.pinv(_build_mel_basis())
-  return np.maximum(1e-10, np.dot(_inv_mel_basis, mel_spectrogram))
-
 def _build_mel_basis():
   n_fft = (hparams.num_freq - 1) * 2
   return librosa.filters.mel(hparams.sample_rate, n_fft, n_mels=hparams.num_mels)