UPDATE June 2023: An extension of this work has also been submitted to IEEE TASLP
This is the official code repo for ASGAN, which goes together with the conference paper:
GAN YOU HEAR ME? RECLAIMING UNCONDITIONAL SPEECH SYNTHESIS FROM DIFFUSION MODELS
Links:
- conference paper arxiv: https://arxiv.org/abs/2210.05271
- audio sample demo page: https://rf5.github.io/slt2022-asgan-demo/
- SLT 2022 proceedings: https://ieeexplore.ieee.org/document/10023153
- quickstart: see below
Figure: the architecture of the mel-spectrogram variant of ASGAN, as given in the paper.
You can load the checkpoint of the best model from the paper (HuBERT variant of ASGAN) using torch hub, so no need to clone the repo!
Simply ensure that all libraries in requirements.txt
are installed, and then:
import torch
from torch import Tensor
model = torch.hub.load('RF5/simple-asgan', 'asgan_hubert_sc09_6')
model = model.eval()
# The below returns a batch of (4, 16000) one second waveforms
# that you can directly save as .wav files.
audio = model.unconditional_generate(4)
The generator nn.Module
is stored in model.g
and dimensions of w
and z
latent variables are stored in model.z_dim
and model.w_dim
.
The model has four convenience functions:
def unconditional_generate(self, N: int) -> Tensor:
""" Generate `N` unconditional audio samples, returning a tensor of shape (N, 16000) """
def generate_from_latent(self, z: Tensor) -> Tensor:
""" Generate waveforms (N, 16000) from latent standard normal `z` (N, z_dim) """
def z2w(self, z: Tensor) -> Tensor:
""" Generate latent w vectors (N, w_dim) from latent standard normal `z` (N, z_dim) """
def generate_from_w(self, w: Tensor) -> Tensor:
""" Generate waveforms (N, 16000) from W latent space `w` (N, w_dim) """
Simple!
For training you must also install deepspeed
.
To prepare the data, use the HuBERT extractor provided in the hubconf.py
.
Put simply:
-
Download the Google Speech Commands dataset
-
Extract the HuBERT
base
layer 6 features for each waveform and save them to some caching directory:import torch import torchaudio hubert = torch.hub.load('RF5/simple-asgan', 'hubert_base') wav, sr = torchaudio.load('/path/to/google_speech_commands_utterance.wav') feats = hubert.get_feats_batched(wav) # (bs, seq_len, dim) torch.save(feats, '/path/to/hubert_feature_cache/uttr_hubert_feats.pt')
Note: if you are using this to preprocess data for training ASGAN, then you should save each utterance's feature tensor as
(seq_len, dim)
. So make sure to squeeze out the first singleton dimension before saving it. -
Use the train, test, and validation splits specified by the google speech commands dataset. I also provide a script to construct this for the SC09 dataset with
split_data.py
:python --root_path /path/to/sc09/ --sc09_only True
. This will save train, validation, and test.csv
files tosplits/
. -
Set the
train_root
indensity/config
to the root directory of the saved hubert feature cache. This should have the same format as the Google Speech Commands dataset.
Simply set the config you wish in config.py
and then you can run the training script with:
python train_asgan.py model=rp_w train_root=/path/to/hubert_feature_cache/ n_valid=400 data_type=hubert_L6 checkpoint_path=./density/runs/cool_training_run/ z_dim=512 rp_w_cfg.z_dim=512 rp_w_cfg.w_layers=1 batch_size=16 lr=2e-3 grad_clip=10 aug_init_p=0.2 stdout_interval=100 validation_interval=2500 n_epochs=800 c_dim=768 rp_w_cfg.c_dim=768 d_lr_mult=0.1 fp16=True preload=False num_workers=12 betas=[0,0.99] rp_w_cfg.equalized_lr=True rp_w_cfg.use_sg3_ff=True rp_w_cfg.D_kernel_size=5 rp_w_cfg.D_block_repeats=[3,3,3,3] use_sc09_splits=True sc09_train_csv=./splits/train.csv sc09_valid_csv=./splits/valid.csv rp_w_cfg.r1_gamma=0.1
Logs will be saved in the checkpoint path. Feel free to tune the training hyperparameters as you see fit.
NOTE: the validation set metrics computed in the training script are quick approximations of the final metrics used in the paper. E.g. FID is computed using HuBERT features on the validation set in the training script, instead of the ResNeXT classifier. This is done to save time and not perform ResNeXT inference on the full validation set output every validation cycle. To obtain the full test set metrics, an evaluation script should be used that follows the method specified in the paper, instead of the approximations made in the training script.
The repository is organized as follows:
├── density
│ ├── augment.py # ADA and update skipping
│ ├── config.py # hyperparameters
│ ├── dataset.py # data loading and processing
│ ├── __init__.py
│ ├── losses.py # training loss
│ ├── metrics.py # logging and evaluation metrics
│ └── models.py # model definition
├── hubconf.py # torchhub integration
├── hubert_feature_reader.py # fairseq hubert feature extractor
├── README.md
├── requirements.txt
├── split_data.py # splits data into train/valid/test subsets
└── train_asgan.py # main training script
Parts of code for this project are adapted from the following repositories -- please make sure to check them out! Thank you to the authors of:
- https://github.com/facebookresearch/fairseq/
- https://github.com/jik876/hifi-gan
- https://github.com/bshall/hifigan
- https://github.com/NVlabs/stylegan3
For the SLT proceedings:
@inproceedings{baas2022asgan,
title={{GAN} you hear me? Reclaiming unconditional speech synthesis from diffusion models},
author={Baas, Matthew and Kamper, Herman},
booktitle={IEEE SLT},
year=2022
}