# Multi-Scale Sub-Band Constant-Q Transform Discriminator for High-Fedility Vocoder [![arXiv](https://img.shields.io/badge/arXiv-Paper-.svg)](https://arxiv.org/abs/2311.14957) [![demo](https://img.shields.io/badge/Vocoder-Demo-red)](https://vocodexelysium.github.io/MS-SB-CQTD/) [![hf](https://img.shields.io/badge/%F0%9F%A4%97%20HuggingFace-Models-pink)](https://huggingface.co./amphion/hifigan_speech_bigdata)

This is the official implementation of the paper "[Multi-Scale Sub-Band Constant-Q Transform Discriminator for High-Fidelity Vocoder](https://arxiv.org/abs/2311.14957)". In this recipe, we will illustrate how to train a high quality HiFi-GAN on LibriTTS, VCTK and LJSpeech via utilizing multiple Time-Frequency-Representation-based Discriminators. There are four stages in total: 1. Data preparation 2. Feature extraction 3. Training 4. Inference > **NOTE:** You need to run every command of this recipe in the `Amphion` root path: > ```bash > cd Amphion > ``` ## 1. Data Preparation ### Dataset Download By default, we utilize the three datasets for training: LibriTTS, VCTK and LJSpeech. How to download them is detailed in [here](../../../datasets/README.md). ### Configuration Specify the dataset path in `exp_config.json`. Note that you can change the `dataset` list to use your preferred datasets. ```json "dataset": [ "ljspeech", "vctk", "libritts", ], "dataset_path": { // TODO: Fill in your dataset path "ljspeech": "[LJSpeech dataset path]", "vctk": "[VCTK dataset path]", "libritts": "[LibriTTS dataset path]", }, ``` ## 2. Features Extraction For HiFiGAN, only the Mel-Spectrogram and the Output Audio are needed for training. ### Configuration Specify the dataset path and the output path for saving the processed data and the training model in `exp_config.json`: ```json // TODO: Fill in the output log path. The default value is "Amphion/ckpts/vocoder" "log_dir": "ckpts/vocoder", "preprocess": { // TODO: Fill in the output data path. The default value is "Amphion/data" "processed_dir": "data", ... }, ``` ### Run Run the `run.sh` as the preproces stage (set `--stage 1`). ```bash sh egs/vocoder/gan/tfr_enhanced_hifigan/run.sh --stage 1 ``` > **NOTE:** The `CUDA_VISIBLE_DEVICES` is set as `"0"` in default. You can change it when running `run.sh` by specifying such as `--gpu "1"`. ## 3. Training ### Configuration We provide the default hyparameters in the `exp_config.json`. They can work on single NVIDIA-24g GPU. You can adjust them based on you GPU machines. ```json "train": { "batch_size": 32, ... } ``` ### Run Run the `run.sh` as the training stage (set `--stage 2`). Specify a experimental name to run the following command. The tensorboard logs and checkpoints will be saved in `Amphion/ckpts/vocoder/[YourExptName]`. ```bash sh egs/vocoder/gan/tfr_enhanced_hifigan/run.sh --stage 2 --name [YourExptName] ``` > **NOTE:** The `CUDA_VISIBLE_DEVICES` is set as `"0"` in default. You can change it when running `run.sh` by specifying such as `--gpu "0,1,2,3"`. If you want to resume or finetune from a pretrained model, run: ```bash sh egs/vocoder/gan/tfr_enhanced_hifigan/run.sh --stage 2 \ --name [YourExptName] \ --resume_type ["resume" for resuming training and "finetune" for loading parameters only] \ --checkpoint Amphion/ckpts/vocoder/[YourExptName]/checkpoint \ ``` > **NOTE:** For multi-gpu training, the `main_process_port` is set as `29500` in default. You can change it when running `run.sh` by specifying such as `--main_process_port 29501`. ## 4. Inference ### Pretrained Vocoder Download We trained a HiFiGAN checkpoint with around 685 hours Speech data. The final pretrained checkpoint is released [here](../../../../pretrained/hifigan/README.md). ### Run Run the `run.sh` as the training stage (set `--stage 3`), we provide three different inference modes, including `infer_from_dataset`, `infer_from_feature`, `and infer_from audio`. ```bash sh egs/vocoder/gan/tfr_enhanced_hifigan/run.sh --stage 3 \ --infer_mode [Your chosen inference mode] \ --infer_datasets [Datasets you want to inference, needed when infer_from_dataset] \ --infer_feature_dir [Your path to your predicted acoustic features, needed when infer_from_feature] \ --infer_audio_dir [Your path to your audio files, needed when infer_form_audio] \ --infer_expt_dir Amphion/ckpts/vocoder/[YourExptName] \ --infer_output_dir Amphion/ckpts/vocoder/[YourExptName]/result \ ``` #### a. Inference from Dataset Run the `run.sh` with specified datasets, here is an example. ```bash sh egs/vocoder/gan/tfr_enhanced_hifigan/run.sh --stage 3 \ --infer_mode infer_from_dataset \ --infer_datasets "libritts vctk ljspeech" \ --infer_expt_dir Amphion/ckpts/vocoder/[YourExptName] \ --infer_output_dir Amphion/ckpts/vocoder/[YourExptName]/result \ ``` #### b. Inference from Features If you want to inference from your generated acoustic features, you should first prepare your acoustic features into the following structure: ```plaintext ┣ {infer_feature_dir} ┃ ┣ mels ┃ ┃ ┣ sample1.npy ┃ ┃ ┣ sample2.npy ``` Then run the `run.sh` with specificed folder direction, here is an example. ```bash sh egs/vocoder/gan/tfr_enhanced_hifigan/run.sh --stage 3 \ --infer_mode infer_from_feature \ --infer_feature_dir [Your path to your predicted acoustic features] \ --infer_expt_dir Amphion/ckpts/vocoder/[YourExptName] \ --infer_output_dir Amphion/ckpts/vocoder/[YourExptName]/result \ ``` #### c. Inference from Audios If you want to inference from audios for quick analysis synthesis, you should first prepare your audios into the following structure: ```plaintext ┣ audios ┃ ┣ sample1.wav ┃ ┣ sample2.wav ``` Then run the `run.sh` with specificed folder direction, here is an example. ```bash sh egs/vocoder/gan/tfr_enhanced_hifigan/run.sh --stage 3 \ --infer_mode infer_from_audio \ --infer_audio_dir [Your path to your audio files] \ --infer_expt_dir Amphion/ckpts/vocoder/[YourExptName] \ --infer_output_dir Amphion/ckpts/vocoder/[YourExptName]/result \ ``` ## Citations ```bibtex @misc{gu2023cqt, title={Multi-Scale Sub-Band Constant-Q Transform Discriminator for High-Fidelity Vocoder}, author={Yicheng Gu and Xueyao Zhang and Liumeng Xue and Zhizheng Wu}, year={2023}, eprint={2311.14957}, archivePrefix={arXiv}, primaryClass={cs.SD} } ```