SaMoye-SVC

expand_train_jxb This branch extends the original model through additional training using data from approximately 4,000 Chinese singers and speakers, and provides instructions for downloading and placing the key weights.

Quick Start: Simple Voice Cloning

This guide covers the fastest way to run the SaMoye SVC model for voice cloning. For this simple inference task, you only need the code located in the SaMoye-SVC/SaMoye-Model/ directory.

Step 1: Download the Pre-trained Model

First, download the required model weights from the link below and place the .pt file inside the SaMoye-SVC/SaMoye-Model/ directory.

• Download Link: https://huggingface.co/karl-wang/SaMoyeSVC/blob/main/sovits_spk_1700h_0020.pt

Step 2: Run the Inference Command

Navigate into the SaMoye-SVC/SaMoye-Model/ directory and execute the following command in your terminal:

python svc_inference.py --config configs/sovits_spk_1700h.yaml --model sovits_spk_1700h_0020.pt --spk path/to/your/speaker.wav --wave path/to/your/content.wav

Important:

You must customize the paths for --spk and --wave.

--spk: This is the path to the audio file of the target voice you want to clone.

--wave: This is the path to the content audio you want to convert.

That's it! It's that simple. You can safely ignore all other folders, as they contain supplementary research materials not required to run the model.

https://github.com/CarlWangChina/SaMoye-SVC/blob/main/SaMoye-SVC-问题回答整理.docx

https://github.com/CarlWangChina/SaMoye-SVC/blob/main/SaMoye-SVC-Question%26Answer.docx

Regarding problems encountered while running the model code, they are documented in the Q&A document available on our homepage. We've prepared two versions: one in Chinese and another in English. If you encounter any bugs in your code, please first consult these documents.

Even dogs can sing a song.

The code is as shown in the repository, and the weight checkpoint can be downloaded from the following: I have uploaded the model weights here on Hugging Face: https://huggingface.co/karl-wang/SaMoyeSVC/tree/main Baidu Netdisk link: https://pan.baidu.com/s/1AxnLlmCSPaMAkEBwUyaI2g?pwd=9999

In the zero-shot scenario of the SaMoye model, we've used 'cat' and 'dog' as reference audios, as well as 'man' and 'woman'. The results generated have all appeared on the homepage.

https://github.com/CarlWangChina/SaMoye-SVC/PetVocolia-Demo Each audio file starts with the target timbre, followed by the generated portion. Please do not skip to the middle to listen to the audio; you will miss the beginning reference audio.

https://github.com/CarlWangChina/SaMoye-SVC/blob/main/test.ipynb

The purpose of test.ipynb is to serve as a script for organizing and processing test data. It is used to handle the test data by generating the necessary files or directories for the inference process. In this context, test.ipynb is a Jupyter Notebook that likely contains instructions or code for preparing test data in the correct format, such as converting or loading data into the appropriate structure that the svc_inference.py script expects.

Script

This model's code is modified and improved based on whisper-vits-svc.

cd model

Training:

python svc_trainer.py -c configs/sovits_spk_1700h.yaml -n sovits_spk_1700h

Tensorboard:

tensorboard --logdir=logs/sovits_spk_1700h --port 12345

Infer script:

Download the ckpt from this link.

python svc_inference.py --config configs/sovits_spk_1700h.yaml --model sovits_spk_1700h_0020.pt --spk spk.wav --wave content.wav

Model Checkpoints Migration Notice

To comply with GitHub's file size limitations and optimize repository structure, all large model checkpoint files (e.g., .pth, .pt, .npy) have been migrated to Hugging Face Hub for centralized storage and management.

Migration Reason

• GitHub restricts individual file sizes to ≤100MB and large files (>2GB) require special handling.
• Hugging Face provides dedicated Large File Storage (LFS) support for model checkpoints.

Hugging Face Repository

All model files are hosted in the following repository:
https://huggingface.co/karl-wang/SaMoyeSVC/tree/main
Specific checkpoint directory:
https://huggingface.co/karl-wang/SaMoyeSVC/tree/main/checkpoints-for-samoye-experiments

Usage Guide

1. Access the Repository:
   Visit the Hugging Face link above and navigate to checkpoints-for-samoye-experiments.

2. Download Checkpoints:
   Select the required model files (see list below) and download them locally.

3. File Placement:
   Place the downloaded files into the project's experiments/ directory to ensure proper code execution.

Migrated File List

The following files have been moved from this repository to Hugging Face: https://huggingface.co/karl-wang/SaMoyeSVC/tree/main

• 3025_nanzhong_00057_005.ppg.npy
• best_model.pth.tar
• hubert-soft-0d54a1f4.pt
• kmeans_10000.pt
• large-v2.pt
• mix2spk92_sing_100k1.pth
• mix2spk92_sing_121k1.pth
• mix2spk92_sing_13k8.pth
• mix2spk92_sing_95k2.pth
• pretrained_model_50.pth
• selfrecord50_sing_100k.pth
• selfrecord50_sing_105k2.pth
• selfrecord50_sing_13k8.pth
• selfrecord50_sing_90k.pth
• sovits5.0_pretrain.pth
• sovitsfrompretrainedOrigin170_100.pth
• toymodel0621spk42_sing_120k.pth
• toymodel0621spk42_sing_130k.pth
• toymodel0621spk42_sing_138k.pth
• trainKMeans10k_10.pth
• trainKMeans10k_25.pth
• trainKMeans10k_50.pth
• trainKMeans10kNOPPG_726k.pth
• trainKMeans900_10.pth
• trainKMeans900_25.pth
• trainKMeans900_50.pth
• trainRVQIDXNOPPG_100epoch.pth
• trainRVQIDXNOPPG_50epoch.pth
• trainRVQNOPPG_71k.pth
• trainRVQNOPPG_91k.pth
• trainTNohubertsoft_50.pth

Large File and Dataset Migration Notice

To keep this GitHub repository lean and efficient, all large files, datasets, experimental configurations, and model weights have been migrated to Hugging Face Hub.

All required assets can be downloaded from our official Hugging Face repository: https://huggingface.co/datasets/karl-wang/SaMoyeSVC/tree/main

Migrated Content Overview

The migrated content includes the following:

1. Archived Experiment Directories: The following complete directories have been archived. Please download the corresponding .zip files and extract them to their original location.

   • experiments/configs
   • experiments/files
   • experiments/train_shpanxin
   • experiments/yongshengTestData
   • experiments/ExperimentResult_20240804_201310

2. Specific Archived Packages from SaMoye-2: These specific large packages have been archived. Please download and extract them to the original paths listed below.

   • raw_data.zip (original path: Singing-Data-Auto-Labeling-and-Diffsinger/raw_data)
   • models.zip (original path: Whisper-SoVITS-Finetune_Contra-Branch-Contrastive-Learning_Rebuild-Branch-Refined/models)
   • singer.zip (original path: Legacy-Diffsinger-Data-Preprocessing-Training-Code/singer)

3. All Other Large Files from SaMoye-2: Additionally, all individual files larger than 5 MB from the SaMoye-2 directory have been moved. This includes audio samples (.wav, .mp3, .aac), model weights (.jit, .npy), dictionaries (.txt), and other auxiliary data.

Important: These files are essential for running experiments, training models, or reproducing results. They are not included in this GitHub repository, so please ensure you download and restore them to their correct original paths to ensure full functionality.

Citation

If you use this work in your research, please cite our paper:

@inproceedings{ijcai2025p1135,
  title     = {AI-Assisted Human-Pet Artistic Musical Co-Creation for Wellness Therapy},
  author    = {Wang, Zihao and Ma, Le and Jin, Yuhang and Feng, Yongsheng and Pan, Xin and Ji, Shulei and Zhang, Kejun},
  booktitle = {Proceedings of the Thirty-Fourth International Joint Conference on
               Artificial Intelligence, {IJCAI-25}},
  publisher = {International Joint Conferences on Artificial Intelligence Organization},
  editor    = {James Kwok},
  pages     = {10216--10224},
  year      = {2025},
  month     = {8},
  note      = {AI, Arts & Creativity},
  doi       = {10.24963/ijcai.2025/1135},
  url       = {https://doi.org/10.24963/ijcai.2025/1135},
}