TCSinger: Zero-Shot Singing Voice Synthesis with Style Transfer and Multi-Level Style Control

Yu Zhang, Ziyue Jiang, Ruiqi Li, Changhao Pan, Jinzheng He, Rongjie Huang, Chuxin Wang, Zhou Zhao | Zhejiang University

PyTorch Implementation of TCSinger (EMNLP 2024): Zero-Shot Singing Voice Synthesis with Style Transfer and Multi-Level Style Control.

We provide our implementation and pre-trained models in this repository.

Visit our demo page for audio samples.

News

• 2024.11: We released the code of TCSinger!
• 2024.09: We released the full dataset of GTSinger!
• 2024.09: TCSinger is accepted by EMNLP 2024!

Key Features

• We present TCSinger, the first zero-shot SVS model for style transfer across cross-lingual speech and singing styles, along with multi-level style control. TCSinger excels in personalized and controllable SVS tasks.
• We introduce the clustering style encoder to extract styles, and the Style and Duration Language Model (S&D-LM) to predict both style information and phoneme duration, addressing style modeling, transfer, and control.
• We propose the style adaptive decoder to generate intricately detailed songs using a novel mel-style adaptive normalization method.
• Experimental results show that TCSinger surpasses baseline models in synthesis quality, singer similarity, and style controllability across various tasks: zero-shot style transfer, multi-level style control, cross-lingual style transfer, and speech-to-singing style transfer.

Quick Start

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

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.

Pre-trained Models

You can use all pre-trained models we provide here. Notably, this TCSinger checkpoint only supports Chinese and English! You should train your own model based on GTSinger for multilingual style transfer and control! Details of each folder are as follows:

Model	Description
TCSinger	Acousitic model (config)
SAPostnet	Postnet model (config)
SDLM	LM model (config)
HIFI-GAN	Neural Vocoder

Dependencies

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

conda create -n tcsinger python=3.10
conda install --yes --file requirements.txt
conda activate tcsinger

Multi-GPU

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.

Inference for bilingual singing voices

Here we provide a speech synthesis pipeline using TCSinger.

1. Prepare TCSinger, SAPostnet, SDLM: Download and put checkpoint at checkpoints/TCSinger, checkpoints/SAPostnet, checkpoints/SDLM.
2. Prepare HIFI-GAN: Download and put checkpoint at checkpoints/hifigan.
3. Prepare prompt information: Provide a prompt_audio (48k) 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 prompt audio path, prompt ph, prompt note, note_dur, note_type. Input these information in Inference/tcsinger.py. Notably, if you want to use Chinese and English data in GTSinger to infer this checkpoint, refer to phone_set, you have to delete _zh or _en in each ph of GTSinger!
4. Infer with tcsinger:

CUDA_VISIBLE_DEVICES=$GPU python inference/tcsinger.py --config egs/sdlm.yaml  --exp_name checkpoints/SDLM

Generated wav files are saved in infer_out by default.

Train your own model based on GTSinger

Data Preparation

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 for each singing voice) and phone_set.json (all phonemes of your dictionary) in data/processed/tc (Note: we provide metadata.json and phone_set.json in GTSinger, 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/tc), 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. Download the global emotion encoder to emotion_encoder_path (training on Chinese only) or train your own global emotion encoder referring to Emotion Encoder based on emotion annotations in GTSinger.
5. Preprocess Dataset:

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

Training TCSinger

1. Train Main Model:

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

2. Train SAPostnet:

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

3. Train SDLM:

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

Inference with TCSinger

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

Acknowledgements

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

Citations

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

@article{zhang2024tcsinger,
  title={TCSinger: Zero-Shot Singing Voice Synthesis with Style Transfer and Multi-Level Style Control},
  author={Zhang, Yu and Jiang, Ziyue and Li, Ruiqi and Pan, Changhao and He, Jinzheng and Huang, Rongjie and Wang, Chuxin and Zhao, Zhou},
  journal={arXiv preprint arXiv:2409.15977},
  year={2024}
}

Disclaimer

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.