## BigVGAN: A Universal Neural Vocoder with Large-Scale Training
#### Sang-gil Lee, Wei Ping, Boris Ginsburg, Bryan Catanzaro, Sungroh Yoon
[[Paper]](https://arxiv.org/abs/2206.04658) - [[Code]](https://github.com/NVIDIA/BigVGAN) - [[Showcase]](https://bigvgan-demo.github.io/) - [[Project Page]](https://research.nvidia.com/labs/adlr/projects/bigvgan/) - [[Weights]](https://huggingface.co/collections/nvidia/bigvgan-66959df3d97fd7d98d97dc9a) - [[Demo]](https://huggingface.co/spaces/nvidia/BigVGAN)
[![PWC](https://img.shields.io/endpoint.svg?url=https://paperswithcode.com/badge/bigvgan-a-universal-neural-vocoder-with-large/speech-synthesis-on-libritts)](https://paperswithcode.com/sota/speech-synthesis-on-libritts?p=bigvgan-a-universal-neural-vocoder-with-large)
## News
- **Sep 2024 (v2.4):**
- We have updated the pretrained checkpoints trained for 5M steps. This is final release of the BigVGAN-v2 checkpoints.
- **Jul 2024 (v2.3):**
- General refactor and code improvements for improved readability.
- Fully fused CUDA kernel of anti-alised activation (upsampling + activation + downsampling) with inference speed benchmark.
- **Jul 2024 (v2.2):** The repository now includes an interactive local demo using gradio.
- **Jul 2024 (v2.1):** BigVGAN is now integrated with 🤗 Hugging Face Hub with easy access to inference using pretrained checkpoints. We also provide an interactive demo on Hugging Face Spaces.
- **Jul 2024 (v2):** We release BigVGAN-v2 along with pretrained checkpoints. Below are the highlights:
- Custom CUDA kernel for inference: we provide a fused upsampling + activation kernel written in CUDA for accelerated inference speed. Our test shows 1.5 - 3x faster speed on a single A100 GPU.
- Improved discriminator and loss: BigVGAN-v2 is trained using a [multi-scale sub-band CQT discriminator](https://arxiv.org/abs/2311.14957) and a [multi-scale mel spectrogram loss](https://arxiv.org/abs/2306.06546).
- Larger training data: BigVGAN-v2 is trained using datasets containing diverse audio types, including speech in multiple languages, environmental sounds, and instruments.
- We provide pretrained checkpoints of BigVGAN-v2 using diverse audio configurations, supporting up to 44 kHz sampling rate and 512x upsampling ratio.
## Installation
The codebase has been tested on Python `3.10` and PyTorch `2.3.1` conda packages with either `pytorch-cuda=12.1` or `pytorch-cuda=11.8`. Below is an example command to create the conda environment:
```shell
conda create -n bigvgan python=3.10 pytorch torchvision torchaudio pytorch-cuda=12.1 -c pytorch -c nvidia
conda activate bigvgan
```
Clone the repository and install dependencies:
```shell
git clone https://github.com/NVIDIA/BigVGAN
cd BigVGAN
pip install -r requirements.txt
```
## Inference Quickstart using 🤗 Hugging Face Hub
Below example describes how you can use BigVGAN: load the pretrained BigVGAN generator from Hugging Face Hub, compute mel spectrogram from input waveform, and generate synthesized waveform using the mel spectrogram as the model's input.
```python
device = 'cuda'
import torch
import bigvgan
import librosa
from meldataset import get_mel_spectrogram
# instantiate the model. You can optionally set use_cuda_kernel=True for faster inference.
model = bigvgan.BigVGAN.from_pretrained('nvidia/bigvgan_v2_24khz_100band_256x', use_cuda_kernel=False)
# remove weight norm in the model and set to eval mode
model.remove_weight_norm()
model = model.eval().to(device)
# load wav file and compute mel spectrogram
wav_path = '/path/to/your/audio.wav'
wav, sr = librosa.load(wav_path, sr=model.h.sampling_rate, mono=True) # wav is np.ndarray with shape [T_time] and values in [-1, 1]
wav = torch.FloatTensor(wav).unsqueeze(0) # wav is FloatTensor with shape [B(1), T_time]
# compute mel spectrogram from the ground truth audio
mel = get_mel_spectrogram(wav, model.h).to(device) # mel is FloatTensor with shape [B(1), C_mel, T_frame]
# generate waveform from mel
with torch.inference_mode():
wav_gen = model(mel) # wav_gen is FloatTensor with shape [B(1), 1, T_time] and values in [-1, 1]
wav_gen_float = wav_gen.squeeze(0).cpu() # wav_gen is FloatTensor with shape [1, T_time]
# you can convert the generated waveform to 16 bit linear PCM
wav_gen_int16 = (wav_gen_float * 32767.0).numpy().astype('int16') # wav_gen is now np.ndarray with shape [1, T_time] and int16 dtype
```
## Local gradio demo
You can run a local gradio demo using below command:
```python
pip install -r demo/requirements.txt
python demo/app.py
```
## Training
Create symbolic link to the root of the dataset. The codebase uses filelist with the relative path from the dataset. Below are the example commands for LibriTTS dataset:
```shell
cd filelists/LibriTTS && \
ln -s /path/to/your/LibriTTS/train-clean-100 train-clean-100 && \
ln -s /path/to/your/LibriTTS/train-clean-360 train-clean-360 && \
ln -s /path/to/your/LibriTTS/train-other-500 train-other-500 && \
ln -s /path/to/your/LibriTTS/dev-clean dev-clean && \
ln -s /path/to/your/LibriTTS/dev-other dev-other && \
ln -s /path/to/your/LibriTTS/test-clean test-clean && \
ln -s /path/to/your/LibriTTS/test-other test-other && \
cd ../..
```
Train BigVGAN model. Below is an example command for training BigVGAN-v2 using LibriTTS dataset at 24kHz with a full 100-band mel spectrogram as input:
```shell
python train.py \
--config configs/bigvgan_v2_24khz_100band_256x.json \
--input_wavs_dir filelists/LibriTTS \
--input_training_file filelists/LibriTTS/train-full.txt \
--input_validation_file filelists/LibriTTS/val-full.txt \
--list_input_unseen_wavs_dir filelists/LibriTTS filelists/LibriTTS \
--list_input_unseen_validation_file filelists/LibriTTS/dev-clean.txt filelists/LibriTTS/dev-other.txt \
--checkpoint_path exp/bigvgan_v2_24khz_100band_256x
```
## Synthesis
Synthesize from BigVGAN model. Below is an example command for generating audio from the model.
It computes mel spectrograms using wav files from `--input_wavs_dir` and saves the generated audio to `--output_dir`.
```shell
python inference.py \
--checkpoint_file /path/to/your/bigvgan_v2_24khz_100band_256x/bigvgan_generator.pt \
--input_wavs_dir /path/to/your/input_wav \
--output_dir /path/to/your/output_wav
```
`inference_e2e.py` supports synthesis directly from the mel spectrogram saved in `.npy` format, with shapes `[1, channel, frame]` or `[channel, frame]`.
It loads mel spectrograms from `--input_mels_dir` and saves the generated audio to `--output_dir`.
Make sure that the STFT hyperparameters for mel spectrogram are the same as the model, which are defined in `config.json` of the corresponding model.
```shell
python inference_e2e.py \
--checkpoint_file /path/to/your/bigvgan_v2_24khz_100band_256x/bigvgan_generator.pt \
--input_mels_dir /path/to/your/input_mel \
--output_dir /path/to/your/output_wav
```
## Using Custom CUDA Kernel for Synthesis
You can apply the fast CUDA inference kernel by using a parameter `use_cuda_kernel` when instantiating BigVGAN:
```python
generator = BigVGAN(h, use_cuda_kernel=True)
```
You can also pass `--use_cuda_kernel` to `inference.py` and `inference_e2e.py` to enable this feature.
When applied for the first time, it builds the kernel using `nvcc` and `ninja`. If the build succeeds, the kernel is saved to `alias_free_activation/cuda/build` and the model automatically loads the kernel. The codebase has been tested using CUDA `12.1`.
Please make sure that both are installed in your system and `nvcc` installed in your system matches the version your PyTorch build is using.
We recommend running `test_cuda_vs_torch_model.py` first to build and check the correctness of the CUDA kernel. See below example command and its output, where it returns `[Success] test CUDA fused vs. plain torch BigVGAN inference`:
```python
python tests/test_cuda_vs_torch_model.py \
--checkpoint_file /path/to/your/bigvgan_generator.pt
```
```shell
loading plain Pytorch BigVGAN
...
loading CUDA kernel BigVGAN with auto-build
Detected CUDA files, patching ldflags
Emitting ninja build file /path/to/your/BigVGAN/alias_free_activation/cuda/build/build.ninja..
Building extension module anti_alias_activation_cuda...
...
Loading extension module anti_alias_activation_cuda...
...
Loading '/path/to/your/bigvgan_generator.pt'
...
[Success] test CUDA fused vs. plain torch BigVGAN inference
> mean_difference=0.0007238413265440613
...
```
If you see `[Fail] test CUDA fused vs. plain torch BigVGAN inference`, it means that the CUDA kernel inference is incorrect. Please check if `nvcc` installed in your system is compatible with your PyTorch version.
## Pretrained Models
We provide the [pretrained models on Hugging Face Collections](https://huggingface.co/collections/nvidia/bigvgan-66959df3d97fd7d98d97dc9a).
One can download the checkpoints of the generator weight (named `bigvgan_generator.pt`) and its discriminator/optimizer states (named `bigvgan_discriminator_optimizer.pt`) within the listed model repositories.
| Model Name | Sampling Rate | Mel band | fmax | Upsampling Ratio | Params | Dataset | Steps | Fine-Tuned |
|:--------------------------------------------------------------------------------------------------------:|:-------------:|:--------:|:-----:|:----------------:|:------:|:--------------------------:|:-----:|:----------:|
| [bigvgan_v2_44khz_128band_512x](https://huggingface.co/nvidia/bigvgan_v2_44khz_128band_512x) | 44 kHz | 128 | 22050 | 512 | 122M | Large-scale Compilation | 5M | No |
| [bigvgan_v2_44khz_128band_256x](https://huggingface.co/nvidia/bigvgan_v2_44khz_128band_256x) | 44 kHz | 128 | 22050 | 256 | 112M | Large-scale Compilation | 5M | No |
| [bigvgan_v2_24khz_100band_256x](https://huggingface.co/nvidia/bigvgan_v2_24khz_100band_256x) | 24 kHz | 100 | 12000 | 256 | 112M | Large-scale Compilation | 5M | No |
| [bigvgan_v2_22khz_80band_256x](https://huggingface.co/nvidia/bigvgan_v2_22khz_80band_256x) | 22 kHz | 80 | 11025 | 256 | 112M | Large-scale Compilation | 5M | No |
| [bigvgan_v2_22khz_80band_fmax8k_256x](https://huggingface.co/nvidia/bigvgan_v2_22khz_80band_fmax8k_256x) | 22 kHz | 80 | 8000 | 256 | 112M | Large-scale Compilation | 5M | No |
| [bigvgan_24khz_100band](https://huggingface.co/nvidia/bigvgan_24khz_100band) | 24 kHz | 100 | 12000 | 256 | 112M | LibriTTS | 5M | No |
| [bigvgan_base_24khz_100band](https://huggingface.co/nvidia/bigvgan_base_24khz_100band) | 24 kHz | 100 | 12000 | 256 | 14M | LibriTTS | 5M | No |
| [bigvgan_22khz_80band](https://huggingface.co/nvidia/bigvgan_22khz_80band) | 22 kHz | 80 | 8000 | 256 | 112M | LibriTTS + VCTK + LJSpeech | 5M | No |
| [bigvgan_base_22khz_80band](https://huggingface.co/nvidia/bigvgan_base_22khz_80band) | 22 kHz | 80 | 8000 | 256 | 14M | LibriTTS + VCTK + LJSpeech | 5M | No |
The paper results are based on the original 24kHz BigVGAN models (`bigvgan_24khz_100band` and `bigvgan_base_24khz_100band`) trained on LibriTTS dataset.
We also provide 22kHz BigVGAN models with band-limited setup (i.e., fmax=8000) for TTS applications.
Note that the checkpoints use `snakebeta` activation with log scale parameterization, which have the best overall quality.
You can fine-tune the models by:
1. downloading the checkpoints (both the generator weight and its discriminator/optimizer states)
2. resuming training using your audio dataset by specifying `--checkpoint_path` that includes the checkpoints when launching `train.py`
## Training Details of BigVGAN-v2
Comapred to the original BigVGAN, the pretrained checkpoints of BigVGAN-v2 used `batch_size=32` with a longer `segment_size=65536` and are trained using 8 A100 GPUs.
Note that the BigVGAN-v2 `json` config files in `./configs` use `batch_size=4` as default to fit in a single A100 GPU for training. You can fine-tune the models adjusting `batch_size` depending on your GPUs.
When training BigVGAN-v2 from scratch with small batch size, it can potentially encounter the early divergence problem mentioned in the paper. In such case, we recommend lowering the `clip_grad_norm` value (e.g. `100`) for the early training iterations (e.g. 20000 steps) and increase the value to the default `500`.
## Evaluation Results of BigVGAN-v2
Below are the objective results of the 24kHz model (`bigvgan_v2_24khz_100band_256x`) obtained from the LibriTTS `dev` sets. BigVGAN-v2 shows noticeable improvements of the metrics. The model also exhibits reduced perceptual artifacts, especially for non-speech audio.
| Model | Dataset | Steps | PESQ(↑) | M-STFT(↓) | MCD(↓) | Periodicity(↓) | V/UV F1(↑) |
|:----------:|:-----------------------:|:-----:|:---------:|:----------:|:----------:|:--------------:|:----------:|
| BigVGAN | LibriTTS | 1M | 4.027 | 0.7997 | 0.3745 | 0.1018 | 0.9598 |
| BigVGAN | LibriTTS | 5M | 4.256 | 0.7409 | 0.2988 | 0.0809 | 0.9698 |
| BigVGAN-v2 | Large-scale Compilation | 3M | 4.359 | 0.7134 | 0.3060 | 0.0621 | 0.9777 |
| BigVGAN-v2 | Large-scale Compilation | 5M | **4.362** | **0.7026** | **0.2903** | **0.0593** | **0.9793** |
## Speed Benchmark
Below are the speed and VRAM usage benchmark results of BigVGAN from `tests/test_cuda_vs_torch_model.py`, using `bigvgan_v2_24khz_100band_256x` as a reference model.
| GPU | num_mel_frame | use_cuda_kernel | Speed (kHz) | Real-time Factor | VRAM (GB) |
|:--------------------------:|:-------------:|:---------------:|:-----------:|:----------------:|:---------:|
| NVIDIA A100 | 256 | False | 1672.1 | 69.7x | 1.3 |
| | | True | 3916.5 | 163.2x | 1.3 |
| | 2048 | False | 1899.6 | 79.2x | 1.7 |
| | | True | 5330.1 | 222.1x | 1.7 |
| | 16384 | False | 1973.8 | 82.2x | 5.0 |
| | | True | 5761.7 | 240.1x | 4.4 |
| NVIDIA GeForce RTX 3080 | 256 | False | 841.1 | 35.0x | 1.3 |
| | | True | 1598.1 | 66.6x | 1.3 |
| | 2048 | False | 929.9 | 38.7x | 1.7 |
| | | True | 1971.3 | 82.1x | 1.6 |
| | 16384 | False | 943.4 | 39.3x | 5.0 |
| | | True | 2026.5 | 84.4x | 3.9 |
| NVIDIA GeForce RTX 2080 Ti | 256 | False | 515.6 | 21.5x | 1.3 |
| | | True | 811.3 | 33.8x | 1.3 |
| | 2048 | False | 576.5 | 24.0x | 1.7 |
| | | True | 1023.0 | 42.6x | 1.5 |
| | 16384 | False | 589.4 | 24.6x | 5.0 |
| | | True | 1068.1 | 44.5x | 3.2 |
## Acknowledgements
We thank Vijay Anand Korthikanti and Kevin J. Shih for their generous support in implementing the CUDA kernel for inference.
## References
- [HiFi-GAN](https://github.com/jik876/hifi-gan) (for generator and multi-period discriminator)
- [Snake](https://github.com/EdwardDixon/snake) (for periodic activation)
- [Alias-free-torch](https://github.com/junjun3518/alias-free-torch) (for anti-aliasing)
- [Julius](https://github.com/adefossez/julius) (for low-pass filter)
- [UnivNet](https://github.com/mindslab-ai/univnet) (for multi-resolution discriminator)
- [descript-audio-codec](https://github.com/descriptinc/descript-audio-codec) and [vocos](https://github.com/gemelo-ai/vocos) (for multi-band multi-scale STFT discriminator and multi-scale mel spectrogram loss)
- [Amphion](https://github.com/open-mmlab/Amphion) (for multi-scale sub-band CQT discriminator)