Language: English Change to Chinese Version

SPA-SVC: Self-supervised Pitch Augmentation for Singing Voice Conversion

This is an official repository of our work SPA-SVC: Self-supervised Pitch Augmentation for Singing Voice Conversion (Accepted by Interspeech2024).

Audio samples are available on the page. Arxiv paper can be found here https://arxiv.org/abs/2406.05692.

SPA-SVC Architecture

(0) Environment Setups

# Creater conda environment, Python version 3.8.18
conda create -n spa-svc python=3.8.18
# Install requirements, torch version 1.31.1+cu116
pip install -r requirements.txt

# Activate conda environment
conda activate spa-svc

(1) Data Preprocessing

nohup python preprocess.py -c configs/spa-svc.yaml >../preprocess_all.log 2>&1 &         

(2) Model Training

nohup python train_diff_singing_enhance.py -c configs/spa-svc-m.yaml >../spa_svc_m.log 2>&1 & # use MSE cycle loss

nohup python train_diff_singing_enhance.py -c configs/spa-svc.yaml >../spa_svc.log 2>&1 & # use SSIM cycle loss

(3) Model Inference

nohup python main_diff.py \
    -i '/path/to/your/audio/directory/' \
    -diff '/path/to/your/model/directory/model.pt' \
    -o '/path/to/your/output/directory/' \
    -k 12 -id 1 -speedup 'auto' -method 'auto' \
    -kstep 100 --gpu_ids 1 \
    > '/path/to/your/log/directory/diffusion-test.log' 2>&1 &