BDDM: Bilateral Denoising Diffusion Models for Fast and High-Quality Speech Synthesis

Related tags

Deep Learning bddm
Overview

Bilateral Denoising Diffusion Models (BDDMs)

GitHub Stars visitors arXiv demo

This is the official PyTorch implementation of the following paper:

BDDM: BILATERAL DENOISING DIFFUSION MODELS FOR FAST AND HIGH-QUALITY SPEECH SYNTHESIS
Max W. Y. Lam, Jun Wang, Dan Su, Dong Yu

Abstract: Diffusion probabilistic models (DPMs) and their extensions have emerged as competitive generative models yet confront challenges of efficient sampling. We propose a new bilateral denoising diffusion model (BDDM) that parameterizes both the forward and reverse processes with a schedule network and a score network, which can train with a novel bilateral modeling objective. We show that the new surrogate objective can achieve a lower bound of the log marginal likelihood tighter than a conventional surrogate. We also find that BDDM allows inheriting pre-trained score network parameters from any DPMs and consequently enables speedy and stable learning of the schedule network and optimization of a noise schedule for sampling. Our experiments demonstrate that BDDMs can generate high-fidelity audio samples with as few as three sampling steps. Moreover, compared to other state-of-the-art diffusion-based neural vocoders, BDDMs produce comparable or higher quality samples indistinguishable from human speech, notably with only seven sampling steps (143x faster than WaveGrad and 28.6x faster than DiffWave).

Paper: Published at ICLR 2022 on OpenReview

BDDM

This implementation supports model training and audio generation, and also provides the pre-trained models for the benchmark LJSpeech and VCTK dataset.

Visit our demo page for audio samples.

Updates:

  • May 20, 2021: Released our follow-up work FastDiff on GitHub, where we futher optimized the speed-and-quality trade-off.
  • May 10, 2021: Added the experiment configurations and model checkpoints for the VCTK dataset.
  • May 9, 2021: Added the searched noise schedules for the LJSpeech and VCTK datasets.
  • March 20, 2021: Released the PyTorch implementation of BDDM with pre-trained models for the LJSpeech dataset.

Recipes:

  • (Option 1) To train the BDDM scheduling network yourself, you can download the pre-trained score network from philsyn/DiffWave-Vocoder (provided at egs/lj/DiffWave.pkl), and follow the training steps below. (Start from Step I.)
  • (Option 2) To search for noise schedules using BDDM, we provide a pre-trained BDDM for LJSpeech at egs/lj/DiffWave-GALR.pkl and for VCTK at egs/vctk/DiffWave-GALR.pkl . (Start from Step III.)
  • (Option 3) To directly generate samples using BDDM, we provide the searched schedules for LJSpeech at egs/lj/noise_schedules and for VCTK at egs/vctk/noise_schedules (check conf.yml for the respective configurations). (Start from Step IV.)

Getting Started

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

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

Dependencies

Step I. Data Preparation and Configuraion

Download the LJSpeech dataset.

For training, we first need to setup a file conf.yml for configuring the data loader, the score and the schedule networks, the training procedure, the noise scheduling and sampling parameters.

Note: Appropriately modify the paths in "train_data_dir" and "valid_data_dir" for training; and the path in "gen_data_dir" for sampling. All dir paths should be link to a directory that store the waveform audios (in .wav) or the Mel-spectrogram files (in .mel).

Step II. Training a Schedule Network

Suppose that a well-trained score network (theta) is stored at $theta_path, we start by modifying "load": $theta_path in conf.yml.

After modifying the relevant hyperparameters for a schedule network (especially "tau"), we can train the schedule network (f_phi in paper) using:

# Training on device 0
sh train.sh 0 conf.yml

Note: In practice, we found that 10K training steps would be enough to obtain a promising scheduling network. This normally takes no more than half an hour for training with one GPU.

Step III. Searching for Noise Schedules

Given a well-trained BDDM (theta, phi), we can now run the noise scheduling algorithm to find the best schedule (optimizing the trade-off between quality and speed).

First, we set "load" in conf.yml to the path of the trained BDDM.

After setting the maximum number of sampling steps in scheduling ("N"), we run:

# Scheduling on device 0
sh schedule.sh 0 conf.yml

Step IV. Evaluation or Generation

For evaluation, we set "gen_data_dir" in conf.yml to the path of a directory that stores the test set of audios (in .wav).

For generation, we set "gen_data_dir" in conf.yml to the path of a directory that stores the Mel-spectrogram (by default in .mel generated by TacotronSTFT or by our dataset loader bddm/loader/dataset.py).

Then, we run:

# Generation/evaluation on device 0 (only support single-GPU scheduling)
sh generate.sh 0 conf.yml

Acknowledgements

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

Citations

@inproceedings{lam2022bddm,
  title={BDDM: Bilateral Denoising Diffusion Models for Fast and High-Quality Speech Synthesis},
  author={Lam, Max WY and Wang, Jun and Su, Dan and Yu, Dong},
  booktitle={International Conference on Learning Representations},
  year={2022}
}

License

Copyright 2022 Tencent

Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with the License. You may obtain a copy of the License at

http://www.apache.org/licenses/LICENSE-2.0

Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the specific language governing permissions and limitations under the License.

Disclaimer

This is not an officially supported Tencent product.

You might also like...
This is the codebase for Diffusion Models Beat GANS on Image Synthesis.

This is the codebase for Diffusion Models Beat GANS on Image Synthesis.

Codebase for Diffusion Models Beat GANS on Image Synthesis.

Codebase for Diffusion Models Beat GANS on Image Synthesis.

A data annotation pipeline to generate high-quality, large-scale speech datasets with machine pre-labeling and fully manual auditing.
A data annotation pipeline to generate high-quality, large-scale speech datasets with machine pre-labeling and fully manual auditing.

About This repository provides data and code for the paper: Scalable Data Annotation Pipeline for High-Quality Large Speech Datasets Development (subm

PyTorch Implementation of PortaSpeech: Portable and High-Quality Generative Text-to-Speech
PyTorch Implementation of PortaSpeech: Portable and High-Quality Generative Text-to-Speech

PortaSpeech - PyTorch Implementation PyTorch Implementation of PortaSpeech: Portable and High-Quality Generative Text-to-Speech. Model Size Module Nor

Facestar dataset. High quality audio-visual recordings of human conversational speech.
Facestar dataset. High quality audio-visual recordings of human conversational speech.

Facestar Dataset Description Existing audio-visual datasets for human speech are either captured in a clean, controlled environment but contain only a

ERISHA is a mulitilingual multispeaker expressive speech synthesis framework. It can transfer the expressivity to the speaker's voice for which no expressive speech corpus is available.
ERISHA is a mulitilingual multispeaker expressive speech synthesis framework. It can transfer the expressivity to the speaker's voice for which no expressive speech corpus is available.

ERISHA: Multilingual Multispeaker Expressive Text-to-Speech Library ERISHA is a multilingual multispeaker expressive speech synthesis framework. It ca

HiFi-GAN: Generative Adversarial Networks for Efficient and High Fidelity Speech Synthesis
HiFi-GAN: Generative Adversarial Networks for Efficient and High Fidelity Speech Synthesis

HiFi-GAN: Generative Adversarial Networks for Efficient and High Fidelity Speech Synthesis Jungil Kong, Jaehyeon Kim, Jaekyoung Bae In our paper, we p

Semantic Segmentation for Real Point Cloud Scenes via Bilateral Augmentation and Adaptive Fusion (CVPR 2021)
Semantic Segmentation for Real Point Cloud Scenes via Bilateral Augmentation and Adaptive Fusion (CVPR 2021)

Semantic Segmentation for Real Point Cloud Scenes via Bilateral Augmentation and Adaptive Fusion (CVPR 2021) This repository is for BAAF-Net introduce

Practical Blind Denoising via Swin-Conv-UNet and Data Synthesis
Practical Blind Denoising via Swin-Conv-UNet and Data Synthesis

Practical Blind Denoising via Swin-Conv-UNet and Data Synthesis [Paper] [Online Demo] The following results are obtained by our SCUNet with purely syn

Comments
  • .mel Data and Scoring Network

    .mel Data and Scoring Network

    Hello fellas,

    im currently working on a project and try to make your model working.

    My first question or issue is your usage of .mel files. Im not having any clue how they get generated. Or from what. My first guess would be sth like Tacotron 2 or sth like that. Since now i was always working with .npy files and im having a really hard time with .mel. Could you tell me where they come from and how they get generated? Would be enough for me to know with with extension they get saved to reproduce that for my data.

    Second one is the Score Network. Do you have any references how you train this and where this comes from?

    Best regards

    LaughingCoffing

    opened by LaughingC0ffin 3
  • LJSpeech data split

    LJSpeech data split

    Hello, thank you for making your code publicly available!

    https://github.com/tencent-ailab/bddm/blob/8c3f807e84f0ebf1a4942a990f369a92cba79c61/egs/lj/conf.yml#L56-L57

    I have a question about the data split. How did you split the entire LJSpeech dataset into train/validation sets in your experiment? The original LJSpeech does not have separate train/dev/test sets, but I would like to follow the procedures of your experiment. I'd appreciate it if you would let me know.

    Best regards.

    opened by yahshibu 2
  • Any way to reproduce your Table 6

    Any way to reproduce your Table 6

    Hi, This is very interesting work! I find that you report pretty good FID results on CIFAR10. I want to know if this repo is the code of audio generation only or if it can be used to reproduce Table 6, too.

    opened by CS123n 1
  • add web demo/model to Huggingface

    add web demo/model to Huggingface

    Hi, would you be interested in adding bddm to Hugging Face? The Hub offers free hosting, and it would make your work more accessible and visible to the rest of the ML community.

    Example from other organizations: Keras: https://huggingface.co/keras-io Microsoft: https://huggingface.co/microsoft Facebook: https://huggingface.co/facebook

    Example spaces with repos: github: https://github.com/salesforce/BLIP Spaces: https://huggingface.co/spaces/salesforce/BLIP

    github: https://github.com/facebookresearch/omnivore Spaces: https://huggingface.co/spaces/akhaliq/omnivore

    and here are guides for adding spaces/models/datasets to your org

    How to add a Space: https://huggingface.co/blog/gradio-spaces how to add models: https://huggingface.co/docs/hub/adding-a-model uploading a dataset: https://huggingface.co/docs/datasets/upload_dataset.html

    Please let us know if you would be interested and if you have any questions, we can also help with the technical implementation.

    opened by AK391 1
Owner
Research repositories.
null
Pytorch-diffusion - A basic PyTorch implementation of 'Denoising Diffusion Probabilistic Models'

PyTorch implementation of 'Denoising Diffusion Probabilistic Models' This reposi

Arthur Juliani 76 Jan 7, 2023
PyTorch Implementation of DiffGAN-TTS: High-Fidelity and Efficient Text-to-Speech with Denoising Diffusion GANs

DiffGAN-TTS - PyTorch Implementation PyTorch implementation of DiffGAN-TTS: High

Keon Lee 157 Jan 1, 2023
High-Resolution Image Synthesis with Latent Diffusion Models

Latent Diffusion Models Requirements A suitable conda environment named ldm can be created and activated with: conda env create -f environment.yaml co

CompVis Heidelberg 5.6k Jan 4, 2023
Denoising Diffusion Probabilistic Models

Denoising Diffusion Probabilistic Models This repo contains code for DDPM training. Based on Denoising Diffusion Probabilistic Models, Improved Denois

Alexander Markov 7 Dec 15, 2022
ILVR: Conditioning Method for Denoising Diffusion Probabilistic Models (ICCV 2021 Oral)

ILVR + ADM This is the implementation of ILVR: Conditioning Method for Denoising Diffusion Probabilistic Models (ICCV 2021 Oral). This repository is h

Jooyoung Choi 225 Dec 28, 2022
Implementation of Retrieval-Augmented Denoising Diffusion Probabilistic Models in Pytorch

Retrieval-Augmented Denoising Diffusion Probabilistic Models (wip) Implementation of Retrieval-Augmented Denoising Diffusion Probabilistic Models in P

Phil Wang 55 Jan 1, 2023
HiFi-GAN: High Fidelity Denoising and Dereverberation Based on Speech Deep Features in Adversarial Networks

HiFiGAN Denoiser This is a Unofficial Pytorch implementation of the paper HiFi-GAN: High Fidelity Denoising and Dereverberation Based on Speech Deep F

Rishikesh (ऋषिकेश) 134 Dec 27, 2022
Pre-trained Deep Learning models and demos (high quality and extremely fast)

OpenVINO™ Toolkit - Open Model Zoo repository This repository includes optimized deep learning models and a set of demos to expedite development of hi

OpenVINO Toolkit 3.4k Dec 31, 2022
A denoising diffusion probabilistic model (DDPM) tailored for conditional generation of protein distograms

Denoising Diffusion Probabilistic Model for Proteins Implementation of Denoising Diffusion Probabilistic Model in Pytorch. It is a new approach to gen

Phil Wang 108 Nov 23, 2022