PyTorch Implementation of TechSinger (AAAI 2025): Technique Controllable Multilingual Singing Voice Synthesis via Flow Matching.

We provide our implementation in this repository.

Visit our demo page for audio samples.

• 2025.2: We released the checkpoints of TechSinger!
• 2025.2: We released the code of TechSinger!
• 2024.12: TechSinger is accepted by AAAI 2025!

We provide an example of how you can generate high-fidelity samples using TechSinger.

To try on your own dataset or GTSinger, simply clone this repo in your local machine provided with NVIDIA GPU + CUDA cuDNN and follow the below instructions.

Simply download the models from . Details of each folder are as follows:

Notably, this TechSinger checkpoint only supports Chinese and English! You should train your own model based on GTSinger for multilingual style transfer and control!

A suitable conda environment named techsinger can be created and activated with:

conda create -n techsinger python=3.10
conda activate techsinger
pip install torch==2.1.0 torchvision==0.16.0 torchaudio==2.1.0 --index-url https://download.pytorch.org/whl/cu118
pip install -r requirements.txt

By default, this implementation uses as many GPUs in parallel as returned by torch.cuda.device_count(). You can specify which GPUs to use by setting the CUDA_DEVICES_AVAILABLE environment variable before running the training module.

Here we provide a singing synthesis pipeline using TechSinger.

1. Prepare stage1, stage2: Download and put checkpoint at checkpoints/stage1, checkpoints/stage2.
2. Prepare HIFI-GAN: Download and put checkpoint at checkpoints/hifigan.
3. Prepare Information: Provide singer id and input target ph, target note for each ph, target note_dur for each ph, target note_type for each ph (rest: 1, lyric: 2, slur: 3), and target technique for each ph (control: 0, technique: 1, random: 2). Input these information in Inference/techsinger.py. Notably, if you want to use data from GTSinger to infer this checkpoint, you need to modify the phonemes in metadata.json of GTSinger (delete "_zh" or "_en" from each phoneme) to ensure that all phonemes are included in phone_set!
4. Infer with techsinger:

export PYTHONPATH=.
CUDA_VISIBLE_DEVICES=$GPU python inference/techsinger.py --config egs/stage2.yaml  --exp_name stage2 --reset

5. You can also use cfg_scale to adjust the degree of the technique. The larger the scale value is, the stronger the degree of the technique will be. The default value is 1.0:

CUDA_VISIBLE_DEVICES=$GPU python inference/techsinger.py --config egs/stage2.yaml  --exp_name stage2 --hparams="cfg_scale=2.0"  --reset

Generated wav files are saved in infer_out by default.

1. Prepare your own singing dataset or download GTSinger.
2. Put metadata.json (including ph, word, item_name, ph_durs, wav_fn, singer, ep_pitches, ep_notedurs, ep_types, and techniques for each singing voice) and phone_set.json (all phonemes of your dictionary) in data/processed/tech (Note: GTSinger provides metadata.json and phone_set.json, but you need to change the wav_fn of each wav in metadata.json to your own absolute path).
3. Set processed_data_dir (data/processed/tech), binary_data_dir,valid_prefixes (list of parts of item names, like ["Chinese#ZH-Alto-1#Mixed_Voice_and_Falsetto#一次就好"]), test_prefixes in the config.
4. Preprocess Dataset:

export PYTHONPATH=.
CUDA_VISIBLE_DEVICES=$GPU python data_gen/tts/bin/binarize.py --config egs/stage1.yaml

1. Train Stage1 Model:

CUDA_VISIBLE_DEVICES=$GPU python tasks/run.py --config egs/stage1.yaml  --exp_name Stage1 --reset

2. Train Stage2 Model:

CUDA_VISIBLE_DEVICES=$GPU python tasks/run.py --config egs/stage2.yaml  --exp_name Stage2 --reset

CUDA_VISIBLE_DEVICES=$GPU python tasks/run.py --config egs/stage2.yaml  --exp_name Stage2 --infer

This implementation uses parts of the code from the following Github repos: NATSpeech, TCSinger, GTSinger as described in our code.

If you find this code useful in your research, please cite our work:

@article{guo2025techsinger,
  title={TechSinger: Technique Controllable Multilingual Singing Voice Synthesis via Flow Matching},
  author={Guo, Wenxiang and Zhang, Yu and Pan, Changhao and Huang, Rongjie and Tang, Li and Li, Ruiqi and Hong, Zhiqing and Wang, Yongqi and Zhao, Zhou},
  journal={arXiv preprint arXiv:2502.12572},
  year={2025}
}

Any organization or individual is prohibited from using any technology mentioned in this paper to generate someone's singing without his/her consent, including but not limited to government leaders, political figures, and celebrities. If you do not comply with this item, you could be in violation of copyright laws.