Merge pull request #3 from MycroftAI/tuning

Tuning for initial run

hparams.py

@@ -10,28 +10,30 @@ hparams = tf.contrib.training.HParams(
   # Audio:
   num_mels=80,
   num_freq=1025,
-  sample_rate=20000,
+  min_mel_freq=125,
+  max_mel_freq=7600,
+  sample_rate=22000,
   frame_length_ms=50,
   frame_shift_ms=12.5,
-  preemphasis=0.97,
   min_level_db=-100,
   ref_level_db=20,
 
   # Model:
   # TODO: add more configurable hparams
   outputs_per_step=5,
+  embedding_dim=512,
 
   # Training:
   batch_size=32,
   adam_beta1=0.9,
   adam_beta2=0.999,
-  initial_learning_rate=0.002,
+  initial_learning_rate=0.0015,
-  decay_learning_rate=True,
+  learning_rate_decay_halflife=100000,
-  use_cmudict=False,  # Use CMUDict during training to learn pronunciation of ARPAbet phonemes
+  use_cmudict=True,   # Use CMUDict during training to learn pronunciation of ARPAbet phonemes
 
   # Eval:
   max_iters=200,
-  griffin_lim_iters=60,
+  griffin_lim_iters=50,
   power=1.5,              # Power to raise magnitudes to prior to Griffin-Lim
 )
 

models/tacotron.py

@@ -39,7 +39,7 @@ class Tacotron():
 
       # Embeddings
       embedding_table = tf.get_variable(
-        'embedding', [len(symbols), 256], dtype=tf.float32,
+        'embedding', [len(symbols), hp.embedding_dim], dtype=tf.float32,
         initializer=tf.truncated_normal_initializer(stddev=0.5))
       embedded_inputs = tf.nn.embedding_lookup(embedding_table, inputs)           # [N, T_in, 256]
 
@@ -127,10 +127,8 @@ class Tacotron():
     '''
     with tf.variable_scope('optimizer') as scope:
       hp = self._hparams
-      if hp.decay_learning_rate:
+      self.learning_rate = tf.train.exponential_decay(
-        self.learning_rate = _learning_rate_decay(hp.initial_learning_rate, global_step)
+        hp.initial_learning_rate, global_step, hp.learning_rate_decay_halflife, 0.5)
-      else:
-        self.learning_rate = tf.convert_to_tensor(hp.initial_learning_rate)
       optimizer = tf.train.AdamOptimizer(self.learning_rate, hp.adam_beta1, hp.adam_beta2)
       gradients, variables = zip(*optimizer.compute_gradients(self.loss))
       self.gradients = gradients
@@ -141,10 +139,3 @@ class Tacotron():
       with tf.control_dependencies(tf.get_collection(tf.GraphKeys.UPDATE_OPS)):
         self.optimize = optimizer.apply_gradients(zip(clipped_gradients, variables),
           global_step=global_step)
-
-
-def _learning_rate_decay(init_lr, global_step):
-  # Noam scheme from tensor2tensor:
-  warmup_steps = 4000.0
-  step = tf.cast(global_step + 1, dtype=tf.float32)
-  return init_lr * warmup_steps**0.5 * tf.minimum(step * warmup_steps**-1.5, step**-0.5)

synthesizer.py

@@ -33,7 +33,6 @@ 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(wav, out)

util/audio.py

@@ -16,16 +16,8 @@ def save_wav(wav, path):
   librosa.output.write_wav(path, wav.astype(np.int16), hparams.sample_rate)
 
 
-def preemphasis(x):
-  return signal.lfilter([1, -hparams.preemphasis], [1], x)
-
-
-def inv_preemphasis(x):
-  return signal.lfilter([1], [1, -hparams.preemphasis], x)
-
-
 def spectrogram(y):
-  D = _stft(preemphasis(y))
+  D = _stft(y)
   S = _amp_to_db(np.abs(D)) - hparams.ref_level_db
   return _normalize(S)
 
@@ -33,21 +25,17 @@ def spectrogram(y):
 def inv_spectrogram(spectrogram):
   '''Converts spectrogram to waveform using librosa'''
   S = _db_to_amp(_denormalize(spectrogram) + hparams.ref_level_db)  # Convert back to linear
-  return inv_preemphasis(_griffin_lim(S ** hparams.power))          # Reconstruct phase
+  return _griffin_lim(S ** hparams.power)                           # Reconstruct phase
 
 
 def inv_spectrogram_tensorflow(spectrogram):
-  '''Builds computational graph to convert spectrogram to waveform using TensorFlow.
+  '''Builds computational graph to convert spectrogram to waveform using TensorFlow.'''
-
-  Unlike inv_spectrogram, this does NOT invert the preemphasis. The caller should call
-  inv_preemphasis on the output after running the graph.
-  '''
   S = _db_to_amp_tensorflow(_denormalize_tensorflow(spectrogram) + hparams.ref_level_db)
   return _griffin_lim_tensorflow(tf.pow(S, hparams.power))
 
 
 def melspectrogram(y):
-  D = _stft(preemphasis(y))
+  D = _stft(y)
   S = _amp_to_db(_linear_to_mel(np.abs(D))) - hparams.ref_level_db
   return _normalize(S)
 
@@ -130,7 +118,8 @@ def _linear_to_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)
+  return librosa.filters.mel(hparams.sample_rate, n_fft, n_mels=hparams.num_mels,
+    fmin=hparams.min_mel_freq, fmax=hparams.max_mel_freq)
 
 def _amp_to_db(x):
   return 20 * np.log10(np.maximum(1e-5, x))