BigVGAN: A Universal Neural Vocoder with Large-Scale Training

In this repository, I try to implement BigVGAN (specifically BigVGAN-base model) [Paper] [Demo].

2023-02-18: The official implementation of BigVGAN is released [Link] 😊

BigVGAN vs HiFi-GAN

1. Leaky Relu --> x + (1/a)*sin^2(ax)

2. MRF --> AMP block

• Up --> Low-pass filter --> Snake1D --> Down --> Low-pass filter

3. MSD --> MRD (Univnet discriminator) UnivNet unofficial github

4. max_fre 8000 --> 12,000 for universal vocoder (sr: 24,000)

• We use the sampling rate of 22,050 and linear spectrogram for input speech.

Pre-requisites

0. Pytorch >=3.9 and torchaudio >= 0.9

1. Download datasets

   1. Download the VCTK dataset

mkdir dataset
cd dataset
wget https://datashare.ed.ac.uk/bitstream/handle/10283/3443/VCTK-Corpus-0.92.zip
unzip VCTK-Corpus-0.92.zip -d VCTK-Corpus
cd ..

Training Exmaple

# VCTK
python preprocess.py

python train_bigvgan_vocoder.py -c configs/vctk_bigvgan.json -m bigvgan

2022-10-10 Question for an initialization method of Snake1D activation

1. Original Snake1D paper states that it is important to initialize a proper value for alpha. For BigVGAN, I just wonder which value and initialization method BigVGAN used. I just set it to 1. The official source code of Snake1D is as below:

# official snake function (https://github.com/AdenosHermes/NeurIPS_2020_Snake/blob/main/Snake_Atmospheric_Temperature.ipynb)
class Snake(nn.Module):#learnable a
  def __init__(self):
      super(Snake, self).__init__()
      self.a = nn.Parameter()
      self.first = True
  def forward(self, x):
    if self.first:
        self.first = False
        a = torch.zeros_like(x[0]).normal_(mean=0,std=50).abs()
        self.a = nn.Parameter(a)
    return (x + (torch.sin(self.a * x) ** 2) / self.a)

2. I will change the alpha parameter with initialization soon
3. I tested some values for initialization. std=0.5 is better than others (1, 5, and 50). However, because I just trained the model for 50,000 steps, I'm not sure it is optimal value for neural vocoder.

2022-07-18 Update (fix alpha value to be trainable)

Thanks to @PiotrDabkowski, I have noticed that alpha is not trained.

I fix it by changing list to torch.nn.ParameterList

Please see this issue (#10)

2022-07-07 Update (use torchaudio.resample with rolloff = 0.5)

After some discussion, I only change the rolloff value (0.25 --> 0.5) and use torchaudio without modification.

After Transposed convolution for upsampling, aliasing may occur so it may be reduced by a low-pass filter.

At first, I misunderstand the frequency information of input feature in AMP block should be maintained after up/down sampling so I tried to modify resampling code to maintain the frequency information under original frequency. But this feature is also upsampled by transposed convolution and aliasing may occuer in this process.

Hence, I only change the rolloff value (0.25 --> 0.5) for anti-aliasing which I mis-implemented.

2022-07-06 (Issues in the rolloff of torchaudio resampling)

Thanks to @Yeongtae, I found that there are something wrong in the cutoff function of low-pass filter.

There are two problems in this function.

1. The reference sampling rate of Cutoff is an lower frequency value between origial sampling rate and resampled sampling rate
2. I used wrong rolloff value of 0.25 --> I will change it soon (0.25 --> 0.5)

torchaudio resampling with rolloff=0.25

• (Left) 22,050
• (Middle) Upsampling (44,100) -->Low-pass filter (rolloff=0.25, torchaudio)
• (Right) Upsampling (44,100) -->Low-pass filter (rolloff=0.25, torchaudio) --> Downsampling (22,050) --> Low-pass filter (rolloff=0.25, torchaudio)

torchaudio resampling with rolloff=0.5

• (Left) 22,050
• (Middle) Upsampling (44,100) -->Low-pass filter (rolloff=0.5, torchaudio)
• (Right) Upsampling (44,100) -->Low-pass filter (rolloff=0.5, torchaudio) --> Downsampling (22,050) --> Low-pass filter (rolloff=0.5, torchaudio)

Modified Resampling with rolloff=0.5

• (Left) 22,050
• (Middle) Upsampling (44,100) -->Low-pass filter (rolloff=0.5, modified torchaudio)
• (Right) Upsampling (44,100) -->Low-pass filter (rolloff=0.5, modified torchaudio) --> Downsampling (22,050) --> Low-pass filter (rolloff=0.5, modified torchaudio)

Upsampling with different rolloff values (Modified Ver.)

• (Left) Upsampling (44,100) -->Low-pass filter (rolloff=0.99, modified torchaudio)
• (Middle)Upsampling (44,100) -->Low-pass filter (rolloff=0.5, modified torchaudio)
• (Right) Upsampling (44,100) -->Low-pass filter (rolloff=0.25, modified torchaudio)

I will update the source code of modified resampling tommorrow. Please do not use this repository for a while...

2022-06-13 (VITS with BigVGAN)

• Build Monotonic Alignment Search first (https://github.com/jaywalnut310/vits)

# Monotonic alignment search
Build MAS (https://github.com/jaywalnut310/vits)

python train_vits_with_bigvgan.py -c configs/vctk_bigvgan_vits.json  -m vits_with_bigvgan

2022-06-12

• Current ver has some redundant parts in some modules (e.g., data_utils have some TTS module. Ignore it plz)

Low-pass filter

• pytorch extension ver with scipy??? StarGAN3
• torchaudio ver (0.9 >=) can use Kaiser window and rolloff (cutoff frequency)

Torchaudio tutorial

torchaudio.transforms is much faster than torchaudio.functional when resampling multiple waveforms using the same paramters.

# beta 
m = 2
n = 6
f_h = 0.6/m
A = 2.285*((m*n)/2 - 1)*math.pi*4*f_h+7.95
beta = 0.1102*(A-8.7)
4.663800127934911

• (2022-06-14) Roll-off (https://pytorch.org/audio/main/_modules/torchaudio/functional/functional.html)

# In torchaudio.functional (https://pytorch.org/audio/main/_modules/torchaudio/functional/functional.html)
 base_freq = min(orig_freq, new_freq)
    # This will perform antialiasing filtering by removing the highest frequencies.
    # At first I thought I only needed this when downsampling, but when upsampling
    # you will get edge artifacts without this, as the edge is equivalent to zero padding,
    # which will add high freq artifacts.
 base_freq *= rolloff
 width = math.ceil(lowpass_filter_width * orig_freq / base_freq)
 idx = torch.arange(-width, width + orig_freq, device=device, dtype=idx_dtype)
 for i in range(new_freq):
     t = (-i / new_freq + idx / orig_freq) * base_freq
     t = t.clamp_(-lowpass_filter_width, lowpass_filter_width)
.
.
.

Hence, I think that using rolloff=0.25 may restrict the bandwith under nyquist freq (fs/2).

Somebody... check it again Please 😢

• (2022-06-12) I need to review this module. I used torchaudio to implement up/downsampling with low-pass filter. I used the rolloff value of 0.25 in the T.resample but I'm not sure this value is equal to sr/(2*m) where m is 2.

torchaudio > 0.9 is needed to use the rolloff parameter for anti-aliasing in T.resample

There are some issues of STFT function in pytorch of 3.9 (When using mixed precision, waveform need to be changed to float before stft )

Results

Audio

I train the BigVGAN-base model with batch size of 64 (using two A100 GPU) and an initial learning rate of 2 × 10⁻⁴

(In original paper, BigVGAN (large model) uses batch size of 32 and an initial learning rate of 1 × 10⁻⁴ to avoid an early training collapse)

For the BigVGAN-base model, I have not yet experienced an early training collapse with batch size of 64 and an initial learning rate of 2 × 10⁻⁴

Automatic Mixed Precision (AMP)

The original paper may not use the AMP during training but this implementation includes AMP. Hence, the results may be different in original setting.

For training with AMP, I change the dtype of representation to float for torchaudio resampling (Need to be changed for precise transformation).

Reference

• BigVGAN: https://arxiv.org/abs/2206.04658
• VITS: https://github.com/jaywalnut310/vits
• UnivNET: https://github.com/mindslab-ai/univnet