k-washi/hifi-gan

環境構築

docker-compose up -d

Docker内

poetry install

vscode extentions install

./.devcontainer/vscode_extentions_install_batch.sh 

Preprocess

poetry run python ./src/dataset/preprocess/select_dataset_audio.py \
    --dataset_dir_list /data/jvs_vc/ /data/lecture_vc/ /data/common_voice/ /data/vc_dataset/ \
    --output_dir /data/voc16k \
    --sample_rate 16000 \
    --audio_time_sec_max 15 \
    --spk_time_sec_max 1800

train

python ./pipeline/train/exp0001.py

model convert

pytorch lightning module から、pytorch module に変更

python ./src/helper/plmodule_to_torch.py \
--checkpoint checkpoints/00002/checkpoint-epoch=0426-val_l1_loss=0.0823.ckpt \
--output ./data/avocodo_v1.pth

tesnsorboard

tensorboard --logdir=./logs/ --host=0.0.0.0 --port=18282

Ref

Avocodo: Generative Adversarial Network for Artifact-Free Vocoder

モデル

モデル	description
avocodo_v1.pth	正規化音で学習したボコーダー

gdown https://drive.google.com/u/1/uc?id=1GgLsmDRqVM9JzT7HcUYW8BR0dh5a1IPm -O avocodo_v1.pth