Unofficial PyTorch Implementation of paper. Most of codes are based on Link

0. LibriTTS dataset (train-clean-100 and train-clean-360) is used.
1. Sampling rate is set to 22050Hz (default).

• Clone this repository.
• Install python requirements. Please refer requirements.txt

0. Run

python prepare_align.py --data_path [LibriTTS DATAPATH]

for some preparations. (You can change the sampling rate by adding --resample_rate [SR])

1. Montreal Forced Aligner (MFA) is used to obtain the alignments between the utterances and the phoneme sequences. 1-1. Download MFA following the command in the website. 1-2. Run the below codes

$ conda activate aligner
$ mfa model download acoustic english_mfa
$ mfa align ......LibriTTS/wav22 lexicon.txt english_us_arpa .........LibriTTS/Textgrid

2. Run

python preprocess.py

(Check input&output data paths)

python train.py --data_path [Preprocessed LibriTTS DATAPATH]

0. Mel generation

python synthesize.py --checkpoint_path [CKPT PATH] --ref_audio [REF AUDIO PATH]

1. Waveform generation (Use hifi-gan)

cd hifi-gan
python inference_e2e.py --checkpoint_file [VOCODER CKPT PATH]