FastDiff: A Fast Conditional Diffusion Model for High-Quality Speech Synthesis

Rongjie Huang, Max W. Y. Lam, Jun Wang, Dan Su, Dong Yu, Yi Ren, Zhou Zhao

PyTorch Implementation of FastDiff (IJCAI'22): a conditional diffusion probabilistic model capable of generating high fidelity speech efficiently.

We provide our implementation and pretrained models as open source in this repository.

Visit our demo page for audio samples.

Our follow-up work might also interest you: ProDiff (ACM Multimedia'22) on GitHub

News

• April.22, 2021: FastDiff accepted by IJCAI 2022.
• June.21, 2022: The LJSpeech checkpoint and demo code are provided.
• August.12, 2022: The VCTK/LibriTTS checkpoints are provided.
• August.25, 2022: FastDiff (tacotron) is provided.
• September, 2022: We release follow-up work ProDiff (ACM Multimedia'22) on GitHub, where we futher optimized the speed-and-quality trade-off.

Quick Started

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

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

Support Datasets and Pretrained Models

You can also use pretrained models we provide here. Details of each folder are as in follows:

More supported datasets are coming soon.

Put the checkpoints in checkpoints/$your_experiment_name/model_ckpt_steps_*.ckpt

Dependencies

See requirements in requirement.txt:

• pytorch
• librosa
• NATSpeech

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 text-to-speech synthesis

Using ProDiff

We provide a more efficient and stable pipeline in and GitHub

Using Tacotron

Download LJSpeech checkpoint for neural vocoding of tacotron output here. We provide a demo in egs/demo_tacotron.ipynb.

Using Portaspeech, DiffSpeech, FastSpeech 2

1. Download LJSpeech checkpoint and put it in checkpoint/FastDiff/model_ckpt_steps_*.ckpt
2. Specify the input $text, and an int-type index $model_index to choose the TTS model. 0(Portaspeech, Ren et al), 1(FastSpeech 2, Ren et al), or 2(DiffSpeech, Liu et al).
3. Set N for reverse sampling, which is a trade off between quality and speed.
4. Run the following command.

CUDA_VISIBLE_DEVICES=$GPU python egs/demo_tts.py --N $N --text $text --model $model_index 

Generated wav files are saved in checkpoints/FastDiff/ by default.
Note: For better quality, it's recommended to finetune the FastDiff model.

Inference from wav file

1. Make wavs directory and copy wav files into the directory.
2. Set N for reverse sampling, which is a trade off between quality and speed.
3. Run the following command.

CUDA_VISIBLE_DEVICES=$GPU python tasks/run.py --config $path/to/config  --exp_name $your_experiment_name --infer --hparams='test_input_dir=wavs,N=$N'

Generated wav files are saved in checkpoints/$your_experiment_name/ by default.

Inference for end-to-end speech synthesis

1. Make mels directory and copy generated mel-spectrogram files into the directory.
   You can generate mel-spectrograms using Tacotron2, Glow-TTS and so forth.
2. Set N for reverse sampling, which is a trade off between quality and speed.
3. Run the following command.

CUDA_VISIBLE_DEVICES=$GPU python tasks/run.py --config $path/to/config --exp_name $your_experiment_name --infer --hparams='test_mel_dir=mels,use_wav=False,N=$N'

Generated wav files are saved in checkpoints/$your_experiment_name/ by default.

Note: If you find the output wav noisy, it's likely because of the mel-preprocessing mismatch between the acoustic and vocoder models.

Train your own model

Data Preparation and Configuraion

1. Set raw_data_dir, processed_data_dir, binary_data_dir in the config file. For custom dataset, please specify configurations of audio preprocessing in modules/FastDiff/config/base.yaml
2. Download dataset to raw_data_dir. Note: the dataset structure needs to follow egs/datasets/audio/*/pre_align.py, or you could rewrite pre_align.py according to your dataset
3. Preprocess Dataset

# Preprocess step: unify the file structure.
python data_gen/tts/bin/pre_align.py --config $path/to/config
# Binarization step: Binarize data for fast IO.
CUDA_VISIBLE_DEVICES=$GPU python data_gen/tts/bin/binarize.py --config $path/to/config

We also provide our processed LJSpeech dataset here.

Training the Refinement Network

CUDA_VISIBLE_DEVICES=$GPU python tasks/run.py --config $path/to/config  --exp_name $your_experiment_name --reset

Training the Noise Predictor Network (Optional)

Refer to Bilateral Denoising Diffusion Models (BDDMs).

Noise Scheduling (Optional)

You can use our pre-derived noise schedule in this time, or refer to Bilateral Denoising Diffusion Models (BDDMs).

Inference

CUDA_VISIBLE_DEVICES=$GPU python tasks/run.py --config $path/to/config  --exp_name $your_experiment_name --infer

Acknowledgements

This implementation uses parts of the code from the following Github repos: NATSpeech, Tacotron2, and DiffWave-Vocoder as described in our code.

Citations

If you find this code useful in your research, please consider citing:

@article{huang2022fastdiff,
  title={FastDiff: A Fast Conditional Diffusion Model for High-Quality Speech Synthesis},
  author={Huang, Rongjie and Lam, Max WY and Wang, Jun and Su, Dan and Yu, Dong and Ren, Yi and Zhao, Zhou},
  booktitle = {Proceedings of the Thirty-First International Joint Conference on
               Artificial Intelligence, {IJCAI-22}},
  publisher = {International Joint Conferences on Artificial Intelligence Organization},
  year={2022}
}

Disclaimer

• This is not an officially supported Tencent product.

• Any organization or individual is prohibited from using any technology mentioned in this paper to generate someone's speech 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.