Update

Signed-off-by: begeekmyfriend <[email protected]>

datasets/audio.py

@@ -14,7 +14,7 @@ def save_wav(wav, path, hparams):
 	f1 = 0.5 * 32767 / max(0.01, np.max(np.abs(wav)))
 	f2 = np.sign(wav) * np.power(np.abs(wav), 0.95)
 	wav = f1 * f2
-	wav = signal.convolve(wav, signal.firwin(hparams.num_freq, [hparams.fmin, hparams.fmax], pass_zero=False, fs=hparams.sample_rate))
+	wav = signal.convolve(wav, signal.firwin(256, [hparams.fmin, hparams.fmax], pass_zero=False, fs=hparams.sample_rate))
 	#proposed by @dsmiller
 	wavfile.write(path, hparams.sample_rate, wav.astype(np.int16))
 

hparams.py

@@ -90,7 +90,6 @@
 	mask_encoder = False, #whether to mask encoder padding while computing attention. Set to True for better prosody but slower convergence.
 	mask_decoder = False, #Whether to use loss mask for padded sequences (if False, <stop_token> loss function will not be weighted, else recommended pos_weight = 20)
 	cross_entropy_pos_weight = 1, #Use class weights to reduce the stop token classes imbalance (by adding more penalty on False Negatives (FN)) (1 = disabled)
-	predict_linear = True, #Whether to add a post-processing network to the Tacotron to predict linear spectrograms (True mode Not tested!!)
 	###########################################################################################################################################
 
 	#Tacotron Training

tacotron/synthesize.py

@@ -2,6 +2,7 @@
 import os
 import re
 import time
+import numpy as np
 from time import sleep
 from datasets import audio
 import tensorflow as tf
@@ -61,7 +62,12 @@ def run_eval(args, checkpoint_path, output_dir, hparams, sentences):
 	batch_sentences = [sentences[i: i+hparams.tacotron_synthesis_batch_size] for i in range(0, len(sentences), delta_size)]
 	start = time.time()
 	for i, batch in enumerate(tqdm(batch_sentences)):
-		synth.eval(batch, log_dir)
+		mel_filename = os.path.join(eval_dir, '{:03d}.npy'.format(i))
+		mel = synth.eval(batch)
+		np.save(mel_filename, mel.T, allow_pickle=False)
+		wav = audio.inv_mel_spectrogram(mel.T, hparams)
+		audio.save_wav(wav, os.path.join(eval_dir, '{:03d}.wav'.format(i)), hparams)
+
 	log('\nGenerated total batch of {} in {:.3f} sec'.format(delta_size, time.time() - start))	
 
 	return eval_dir
@@ -87,7 +93,7 @@ def run_synthesis(args, checkpoint_path, output_dir, hparams):
 	with open(metadata_filename, encoding='utf-8') as f:
 		metadata = [line.strip().split('|') for line in f]
 		frame_shift_ms = hparams.hop_size / hparams.sample_rate
-		hours = sum([int(x[4]) for x in metadata]) * frame_shift_ms / (3600)
+		hours = sum([int(x[2]) for x in metadata]) * frame_shift_ms / (3600)
 		log('Loaded metadata for {} examples ({:.2f} hours)'.format(len(metadata), hours))
 
 	metadata = [metadata[i: i+hparams.tacotron_synthesis_batch_size] for i in range(0, len(metadata), hparams.tacotron_synthesis_batch_size)]
@@ -97,13 +103,12 @@ def run_synthesis(args, checkpoint_path, output_dir, hparams):
 	wav_dir = os.path.join(args.input_dir, 'audio')
 	with open(os.path.join(synth_dir, 'map.txt'), 'w') as file:
 		for i, meta in enumerate(tqdm(metadata)):
-			texts = [m[5] for m in meta]
-			mel_filenames = [os.path.join(mel_dir, m[1]) for m in meta]
-			wav_filenames = [os.path.join(wav_dir, m[0]) for m in meta]
+			texts = [m[3] for m in meta]
+			mel_filenames = [os.path.join(mel_dir, m[0]) for m in meta]
 			basenames = [os.path.basename(m).replace('.npy', '').replace('mel-', '') for m in mel_filenames]
 			mel_output_filenames, speaker_ids = synth.synthesize(texts, basenames, synth_dir, None, mel_filenames)
 
-			for elems in zip(wav_filenames, mel_filenames, mel_output_filenames, speaker_ids, texts):
+			for elems in zip(mel_filenames, mel_output_filenames, speaker_ids, texts):
 				file.write('|'.join([str(x) for x in elems]) + '\n')
 	log('synthesized mel spectrograms at {}'.format(synth_dir))
 	return os.path.join(synth_dir, 'map.txt')

tacotron/synthesizer.py

@@ -100,7 +100,7 @@ def synthesize(self, texts, basenames, out_dir, log_dir, mel_filenames):
 
 		return saved_mels_paths, speaker_ids
 
-	def eval(self, batch, out_dir):
+	def eval(self, batch):
 		hparams = self._hparams
 		cleaner_names = [x.strip() for x in hparams.cleaners.split(',')]
 		seqs = [np.asarray(text_to_sequence(text, cleaner_names)) for text in batch]
@@ -120,22 +120,7 @@ def eval(self, batch, out_dir):
 		mels = [mel[:target_length, :] for mel, target_length in zip(mels, target_lengths)]
 		assert len(mels) == len(batch)
 
-		#save wav (mel -> wav)
-		mel_filenames = []
-		for i, mel in enumerate(mels):
-			mel_filename = os.path.join(out_dir, '{:03d}.npy'.format(i))
-			np.save(mel_filename, mel.T, allow_pickle=False)
-			wav = audio.inv_mel_spectrogram(mel.T, self._hparams)
-			audio.save_wav(wav, os.path.join(out_dir, '{:03d}.wav'.format(i)), self._hparams)
-			mel_filenames.append(mel_filename)
-		return mel_filenames
-
-		results = []
-		for i, linear in enumerate(linears):
-			linear_wav = self.session.run(self.linear_wav_outputs, feed_dict={self.linear_spectrograms: linear})
-			wav = audio.inv_preemphasis(linear_wav, hparams.preemphasis)
-			results.append(wav)
-		return np.concatenate(results)
+		return np.concatenate(mels)
 
 	def _round_up(self, x, multiple):
 		remainder = x % multiple