Better mel feature extraction for better alignment

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

common/preprocessor.py

@@ -93,6 +93,7 @@ def _process_utterance(wav_dir, mel_dir, basename, wav_file, text, hparams):
 	# Write the spectrogram and audio to disk
 	filename = f'{basename}.npy'
 	np.save(os.path.join(wav_dir, filename), wav, allow_pickle=False)
+	np.save(os.path.join(mel_dir, filename), mel_spectrogram, allow_pickle=False)
 
 	# Return a tuple describing this training example
 	return (filename, time_steps, mel_frames, text)

scripts/train_tacotron2.sh

@@ -1 +1 @@
-CUDA_VISIBLE_DEVICES=0 python train.py --amp-run -o logs --init-lr 1e-3 --final-lr 1e-5 --epochs 200 -bs 32 --weight-decay 1e-6 --log-file nvlog.json --dataset-path training_data --training-anchor-dirs tts_fanfanli_22050 tts_xiaoya_22050 tts_yangluzhuo_22050 tts_yuanzhonglu_22050
+CUDA_VISIBLE_DEVICES=0 python train.py --amp-run -o logs --init-lr 1e-3 --final-lr 1e-5 --epochs 200 -bs 32 --weight-decay 1e-6 --log-file nvlog.json --dataset-path training_data --training-anchor-dirs --load-mel-from-disk tts_fanfanli_22050 tts_xiaoya_22050 tts_yangluzhuo_22050 tts_yuanzhonglu_22050

tacotron2/loader.py

@@ -40,7 +40,7 @@ def parse_tacotron2_args(parent, add_help=False):
     # misc parameters
     parser.add_argument('--mask-padding', default=False, type=bool, help='Use mask padding')
     parser.add_argument('--n-mel-channels', default=80, type=int, help='Number of bins in mel-spectrograms')
-    parser.add_argument('--mel_pad_val', default=-11.5129, type=float, help='Corresponding to silence')
+    parser.add_argument('--mel_pad_val', default=-5, type=float, help='Corresponding to silence')
 
     # symbols parameters
     global symbols