search
HomeTechnology peripheralsAIWith the help of Schrödinger Bridge, Zhu Jun's team at Tsinghua University develops a new speech synthesis system to address proliferation challenges

Recently, a speech synthesis system based on the Schrödinger Bridge [1] released by Professor Zhu Jun’s research group from the Department of Computer Science at Tsinghua University, relies on its “data-to-data” generation paradigm, defeating both in terms of sample quality and sampling speed. The "noise-to-data" paradigm of diffusion models.

With the help of Schrödinger Bridge, Zhu Juns team at Tsinghua University develops a new speech synthesis system to address proliferation challenges

Paper link: https://arxiv.org/abs/2312.03491
Project website: https://bridge-tts.github.io/
Code implementation: https://github.com/thu-ml/Bridge-TTS

##Problem background

Since 2021, diffusion models have begun to become one of the core generation methods in the field of text-to-speech synthesis (TTS). First, methods such as Grad-TTS [2] proposed by Huawei's Noah's Ark Laboratory and DiffSinger [3] proposed by Zhejiang University have achieved high generation quality. Since then, many research works have effectively improved the sampling speed of diffusion models, such as through prior optimization [2, 3, 4], model distillation [5, 6], residual prediction [7] and other methods. However, as shown in this study, because the diffusion model is limited to the generation paradigm of "noise to data", its prior distribution always provides limited information for the generation target and cannot fully utilize the conditional information.
The latest research work in the field of speech synthesis, Bridge-TTS, relies on its generation framework based on Schrödinger bridge to realize the generation process of "data to data". For the first time, the priori speech synthesis Information is modified from noise to clean data, and is modified from distribution to deterministic representation.

The main architecture of this method is shown in the figure above. The input text is first extracted through the text encoder to extract the latent space representation of the generated target (mel-spectrogram, mel spectrum). Thereafter, unlike the diffusion model that incorporates this information into the noise distribution or uses it as conditional information, the Bridge-TTS method ‍ supports directly using it as prior information and supports random or deterministic sampling, High Quality , quicklygenerate targets.

Work results

In verifying the quality of speech synthesis On the standard data set LJ-Speech, the research team compared Bridge-TTS with 9 high-quality speech synthesis systems and accelerated sampling methods of diffusion models. As shown below, this method beats high-quality diffusion model-based TTS systems [2,3,7] in sample quality (1000 steps, 50 steps sampling) and in sampling speed, without any post-processing such as additional model distillation, it surpasses many acceleration methods, such as residual prediction, progressive distillation, and the latest consistency distillation [5, 6, 7].
With the help of Schrödinger Bridge, Zhu Juns team at Tsinghua University develops a new speech synthesis system to address proliferation challenges
The following are examples of the generation effects of Bridge-TTS and the diffusion model-based method. For more generation sample comparisons, please visit the project website: https://bridge-tts.github. io/

  • 1000 step synthesis effect comparison

Enter text: "Printing, then, for our purpose, may be considered as the art of making books by means of movable types." With the help of Schrödinger Bridge, Zhu Juns team at Tsinghua University develops a new speech synthesis system to address proliferation challenges
  • 4 Step Synthesis Effect Comparison

Input text: "The first books were printed in black letter, i.e. the letter which was a Gothic development of the ancient Roman character,」With the help of Schrödinger Bridge, Zhu Juns team at Tsinghua University develops a new speech synthesis system to address proliferation challenges
  • ## 2 Comparison of step synthesis effects

Enter text: "The prison population fluctuated a great deal,"
With the help of Schrödinger Bridge, Zhu Juns team at Tsinghua University develops a new speech synthesis system to address proliferation challengesThe following shows Bridge- TTS A case of deterministic synthesis (ODE sampling) in steps 2 and 4. In 4-step synthesis, this method significantly synthesizes more sample details than the diffusion model, and there is no problem of residual noise. In a 2-step synthesis, this method exhibits completely pure sampling trajectories, and refines more generated details at each step.
With the help of Schrödinger Bridge, Zhu Juns team at Tsinghua University develops a new speech synthesis system to address proliferation challenges
In the frequency domain, more generated samples are shown below. In 1000 steps of synthesis, this method generates higher quality mel compared to the diffusion model. Spectrum, when the number of sampling steps drops to 50, the diffusion model has sacrificed some sampling details, while the method based on Schrödinger bridge still maintains high-quality generation effects. In 4-step and 2-step synthesis, this method does not require distillation, multi-stage training, and adversarial loss functions, and still achieves high-quality generation effects.

With the help of Schrödinger Bridge, Zhu Juns team at Tsinghua University develops a new speech synthesis system to address proliferation challenges

##Comparison of the mel spectra of Bridge-TTS and the diffusion model-based method in 1000 steps of synthesis

With the help of Schrödinger Bridge, Zhu Juns team at Tsinghua University develops a new speech synthesis system to address proliferation challenges

Comparison of Mel spectra between Bridge-TTS and diffusion model-based methods in 50-step synthesis

With the help of Schrödinger Bridge, Zhu Juns team at Tsinghua University develops a new speech synthesis system to address proliferation challenges

Comparison of Mel spectra between Bridge-TTS and diffusion model-based methods in 4-step synthesis
With the help of Schrödinger Bridge, Zhu Juns team at Tsinghua University develops a new speech synthesis system to address proliferation challenges
Comparison of Mel spectra between Bridge-TTS and diffusion model-based methods in 2-step synthesis

Once Bridge-TTS was released, it attracted enthusiastic attention on Twitter with its novel design and high-quality synthesis effects in speech synthesis. , received more than a hundred retweets and hundreds of likes, was selected into Huggingface’s Daily Paper on 12.7 and won the first place in the support rate on that day, and was also featured in LinkedIn, Weibo, Zhihu, Xiaohongshu and other domestic It was followed and reported on external platforms.

With the help of Schrödinger Bridge, Zhu Juns team at Tsinghua University develops a new speech synthesis system to address proliferation challenges

Many foreign language websites also reported and discussed:

With the help of Schrödinger Bridge, Zhu Juns team at Tsinghua University develops a new speech synthesis system to address proliferation challenges

Method introduction

##Schrodinger Bridge is a type of deep
generative model that has recently emerged following the diffusion model, has preliminary applications in image generation, image translation and other fields [8,9]. Unlike the diffusion model, which establishes a transformation process between data and Gaussian noise, the Schrödinger bridge supports transformation between any two boundary distributions. In the study of Bridge-TTS, the authors proposed a speech synthesis framework based on the Schrödinger bridge between paired data, which flexibly supports a variety of forward processes, prediction targets, and sampling processes. An overview of its method is shown in the figure below:

With the help of Schrödinger Bridge, Zhu Juns team at Tsinghua University develops a new speech synthesis system to address proliferation challenges

  • Forward process: This study combines strong information priors and generation goals A fully solvable Schrödinger bridge is built between them, supporting flexible forward process selection, such as symmetric noise strategy:, constantWith the help of Schrödinger Bridge, Zhu Juns team at Tsinghua University develops a new speech synthesis system to address proliferation challenges, and With the help of Schrödinger Bridge, Zhu Juns team at Tsinghua University develops a new speech synthesis system to address proliferation challengesasymmetric noise strategy: , linear With the help of Schrödinger Bridge, Zhu Juns team at Tsinghua University develops a new speech synthesis system to address proliferation challenges, and With the help of Schrödinger Bridge, Zhu Juns team at Tsinghua University develops a new speech synthesis system to address proliferation challengesvariance-preserving (VP) noise strategies that correspond directly to diffusion models. This method found that in speech synthesis tasks, asymmetric noise strategies: linear (gmax) and VP processes have better generation effects than symmetric noise strategies. With the help of Schrödinger Bridge, Zhu Juns team at Tsinghua University develops a new speech synthesis system to address proliferation challenges

With the help of Schrödinger Bridge, Zhu Juns team at Tsinghua University develops a new speech synthesis system to address proliferation challenges

With the help of Schrödinger Bridge, Zhu Juns team at Tsinghua University develops a new speech synthesis system to address proliferation challenges

  • Model training: This method maintains many advantages of the diffusion model training process, such as single stage, single model, and single loss function. And it compares various methods of model parameterization (Model parameterization), that is, the selection of network training targets, including noise prediction (Noise), generation target prediction (Data), and flow matching technology corresponding to the diffusion model [10,11] Velocity prediction (Velocity), etc. The article found that when the generation target, that is, the mel spectrum, is used as the network prediction target, relatively better generation results can be achieved.

With the help of Schrödinger Bridge, Zhu Juns team at Tsinghua University develops a new speech synthesis system to address proliferation challengesWith the help of Schrödinger Bridge, Zhu Juns team at Tsinghua University develops a new speech synthesis system to address proliferation challenges

  • Sampling process: Thanks to the fact that the Schrödinger bridge is completely solvable in this study By transforming the forward-backward SDE system corresponding to the Schrödinger bridge, the authors obtained Bridge SDE and Bridge ODE for inference. At the same time, due to the slow speed of direct simulation of Bridge SDE/ODE inference, in order to speed up sampling, this study used the exponential integrator commonly used in diffusion models [12,13], and gave the first-order SDE and ODE sampling forms of the Schrödinger bridge:

With the help of Schrödinger Bridge, Zhu Juns team at Tsinghua University develops a new speech synthesis system to address proliferation challenges

#In 1-step sampling, the sampling forms of first-order SDE and ODE jointly degenerate into single-step prediction of the network. At the same time, they are closely related to posterior sampling/diffusion model DDIM sampling, and the article gives a detailed analysis in the appendix. The article also provides the second-order sampling SDE and ODE sampling algorithms of the Schrödinger bridge. The authors found that in speech synthesis, the generation quality is similar to a first-order sampling process.

In other tasks such as speech enhancement, speech separation, speech editing and other tasks where prior information is also strong, the authors expect that this research will also bring greater Value.

About the author

This study has three co-first authors: Chen Zehua, He Guande and Zheng Kaiwen both belong to Zhu Jun’s research group in the Department of Computer Science at Tsinghua University. The corresponding author of the article is Professor Zhu Jun. Tan Xu, chief research manager of Microsoft Research Asia, is a project collaborator.

With the help of Schrödinger Bridge, Zhu Juns team at Tsinghua University develops a new speech synthesis system to address proliferation challenges

#                                                       Tan Xu, Chief Research Manager, Microsoft Research Asia

With the help of Schrödinger Bridge, Zhu Juns team at Tsinghua University develops a new speech synthesis system to address proliferation challenges

Chen Zehua is a Shuimu Scholar postdoctoral fellow in the Department of Computer Science at Tsinghua University. His main research direction is probabilistic generative models and their applications in speech, sound effects, bioelectrical signal synthesis, etc. He has interned at many companies such as Microsoft, JD.com, and TikTok, and published many papers at important international conferences in the field of speech and machine learning, such as ICML/NeurIPS/ICASSP.

With the help of Schrödinger Bridge, Zhu Juns team at Tsinghua University develops a new speech synthesis system to address proliferation challenges

He Guande is a third-year master's student at Tsinghua University. His main research direction is uncertainty estimation and generative models. He has previously participated in conferences such as ICLR and Published the paper as the first author.

With the help of Schrödinger Bridge, Zhu Juns team at Tsinghua University develops a new speech synthesis system to address proliferation challenges

Zheng Kaiwen is a second-year master's student at Tsinghua University. His main research direction is the theory and algorithm of deep generative models, and their applications in image, audio and 3D generation. He has previously published many papers at top conferences such as ICML/NeurIPS/CVPR, involving technologies such as flow matching and exponential integrators in diffusion models.

References:
[1] Zehua Chen, Guande He , Kaiwen Zheng, Xu Tan, and Jun Zhu. Schrodinger Bridges Beat Diffusion Models on Text-to-Speech Synthesis. arXiv preprint arXiv:2312.03491, 2023.
[2] Vadim Popov, Ivan Vovk, Vladimir Gogoryan, Tasnima Sadekova, and Mikhail A. Kudinov. Grad-TTS: A Diffusion Probabilistic Model for Text-to-Speech. In ICML, 2021.
[3] Jinglin Liu, Chengxi Li, Yi Ren, Feiyang Chen, and Zhou Zhao. DiffSinger: Singing Voice Synthesis via Shallow Diffusion Mechanism. In AAAI, 2022.
[4] Sang-gil Lee, Heeseung Kim, Chaehun Shin, Xu Tan, Chang Liu, Qi Meng, Tao Qin, Wei Chen, Sungroh Yoon, and Tie-Yan Liu. PriorGrad : Improving Conditional Denoising Diffusion Models with Data-Dependent Adaptive Prior. In ICLR, 2022.
##[5] Rongjie Huang, Zhou Zhao, Huadai Liu, Jinglin Liu , Chenye Cui, and Yi Ren. ProDiff: Progressive Fast Diffusion Model For High-Quality Text-to-Speech. In ACM Multimedia, 2022.
[6 ] Zhen Ye, Wei Xue, Xu Tan, Jie Chen, Qifeng Liu, and Yike Guo. CoMoSpeech: One-Step Speech and Singing Voice Synthesis via Consistency Model. In ACM Multimedia, 2023.
[7] Zehua Chen, Yihan Wu, Yichong Leng, Jiawei Chen, Haohe Liu, Xu Tan, Yang Cui, Ke Wang, Lei He, Sheng Zhao, Jiang Bian, and Danilo P. Mandic. ResGrad: Residual Denoising Diffusion Probabilistic Models for Text to Speech. arXiv preprint arXiv:2212.14518, 2022.
##[8] Yuyang Shi, Valentin De Bortoli, Andrew Campbell , and Arnaud Doucet. Diffusion Schrödinger Bridge Matching. In NeurIPS 2023.
[9] Guan-Horng Liu, Arash Vahdat, De-An Huang, Evangelos A . Theodorou, Weili Nie, and Anima Anandkumar. I2SB: Image-to-Image Schrödinger Bridge. In ICML, 2023.
##[10] Yaron Lipman, Ricky T. Q. Chen, Heli Ben-Hamu, Maximilian Nickel, and Matt Le. Flow Matching for Generative Modeling. In ICLR, 2023.
##[11] Kaiwen Zheng, Cheng Lu, Jianfei Chen, and Jun Zhu. Improved Techniques for Maximum Likelihood Estimation for Diffusion ODEs. In ICML, 2023.
##[12] Cheng Lu, Yuhao Zhou, Fan Bao, Jianfei Chen, Chongxuan Li, and Jun Zhu. DPM-Solver: A Fast ODE Solver for Diffusion Probabilistic Model Sampling in Around 10 Steps. In NeurIPS, 2022.
[13] Kaiwen Zheng, Cheng Lu, Jianfei Chen, and Jun Zhu. DPM-Solver-v3: Improved Diffusion ODE Solver with Empirical Model Statistics. In NeurIPS, 2023.

The above is the detailed content of With the help of Schrödinger Bridge, Zhu Jun's team at Tsinghua University develops a new speech synthesis system to address proliferation challenges. For more information, please follow other related articles on the PHP Chinese website!

Statement
This article is reproduced at:机器之心. If there is any infringement, please contact admin@php.cn delete
DSA如何弯道超车NVIDIA GPU?DSA如何弯道超车NVIDIA GPU?Sep 20, 2023 pm 06:09 PM

你可能听过以下犀利的观点:1.跟着NVIDIA的技术路线,可能永远也追不上NVIDIA的脚步。2.DSA或许有机会追赶上NVIDIA,但目前的状况是DSA濒临消亡,看不到任何希望另一方面,我们都知道现在大模型正处于风口位置,业界很多人想做大模型芯片,也有很多人想投大模型芯片。但是,大模型芯片的设计关键在哪,大带宽大内存的重要性好像大家都知道,但做出来的芯片跟NVIDIA相比,又有何不同?带着问题,本文尝试给大家一点启发。纯粹以观点为主的文章往往显得形式主义,我们可以通过一个架构的例子来说明Sam

阿里云通义千问14B模型开源!性能超越Llama2等同等尺寸模型阿里云通义千问14B模型开源!性能超越Llama2等同等尺寸模型Sep 25, 2023 pm 10:25 PM

2021年9月25日,阿里云发布了开源项目通义千问140亿参数模型Qwen-14B以及其对话模型Qwen-14B-Chat,并且可以免费商用。Qwen-14B在多个权威评测中表现出色,超过了同等规模的模型,甚至有些指标接近Llama2-70B。此前,阿里云还开源了70亿参数模型Qwen-7B,仅一个多月的时间下载量就突破了100万,成为开源社区的热门项目Qwen-14B是一款支持多种语言的高性能开源模型,相比同类模型使用了更多的高质量数据,整体训练数据超过3万亿Token,使得模型具备更强大的推

ICCV 2023揭晓:ControlNet、SAM等热门论文斩获奖项ICCV 2023揭晓:ControlNet、SAM等热门论文斩获奖项Oct 04, 2023 pm 09:37 PM

在法国巴黎举行了国际计算机视觉大会ICCV(InternationalConferenceonComputerVision)本周开幕作为全球计算机视觉领域顶级的学术会议,ICCV每两年召开一次。ICCV的热度一直以来都与CVPR不相上下,屡创新高在今天的开幕式上,ICCV官方公布了今年的论文数据:本届ICCV共有8068篇投稿,其中有2160篇被接收,录用率为26.8%,略高于上一届ICCV2021的录用率25.9%在论文主题方面,官方也公布了相关数据:多视角和传感器的3D技术热度最高在今天的开

复旦大学团队发布中文智慧法律系统DISC-LawLLM,构建司法评测基准,开源30万微调数据复旦大学团队发布中文智慧法律系统DISC-LawLLM,构建司法评测基准,开源30万微调数据Sep 29, 2023 pm 01:17 PM

随着智慧司法的兴起,智能化方法驱动的智能法律系统有望惠及不同群体。例如,为法律专业人员减轻文书工作,为普通民众提供法律咨询服务,为法学学生提供学习和考试辅导。由于法律知识的独特性和司法任务的多样性,此前的智慧司法研究方面主要着眼于为特定任务设计自动化算法,难以满足对司法领域提供支撑性服务的需求,离应用落地有不小的距离。而大型语言模型(LLMs)在不同的传统任务上展示出强大的能力,为智能法律系统的进一步发展带来希望。近日,复旦大学数据智能与社会计算实验室(FudanDISC)发布大语言模型驱动的中

百度文心一言全面向全社会开放,率先迈出重要一步百度文心一言全面向全社会开放,率先迈出重要一步Aug 31, 2023 pm 01:33 PM

8月31日,文心一言首次向全社会全面开放。用户可以在应用商店下载“文心一言APP”或登录“文心一言官网”(https://yiyan.baidu.com)进行体验据报道,百度计划推出一系列经过全新重构的AI原生应用,以便让用户充分体验生成式AI的理解、生成、逻辑和记忆等四大核心能力今年3月16日,文心一言开启邀测。作为全球大厂中首个发布的生成式AI产品,文心一言的基础模型文心大模型早在2019年就在国内率先发布,近期升级的文心大模型3.5也持续在十余个国内外权威测评中位居第一。李彦宏表示,当文心

致敬TempleOS,有开发者创建了启动Llama 2的操作系统,网友:8G内存老电脑就能跑致敬TempleOS,有开发者创建了启动Llama 2的操作系统,网友:8G内存老电脑就能跑Oct 07, 2023 pm 10:09 PM

不得不说,Llama2的「二创」项目越来越硬核、有趣了。自Meta发布开源大模型Llama2以来,围绕着该模型的「二创」项目便多了起来。此前7月,特斯拉前AI总监、重回OpenAI的AndrejKarpathy利用周末时间,做了一个关于Llama2的有趣项目llama2.c,让用户在PyTorch中训练一个babyLlama2模型,然后使用近500行纯C、无任何依赖性的文件进行推理。今天,在Karpathyllama2.c项目的基础上,又有开发者创建了一个启动Llama2的演示操作系统,以及一个

AI技术在蚂蚁集团保险业务中的应用:革新保险服务,带来全新体验AI技术在蚂蚁集团保险业务中的应用:革新保险服务,带来全新体验Sep 20, 2023 pm 10:45 PM

保险行业对于社会民生和国民经济的重要性不言而喻。作为风险管理工具,保险为人民群众提供保障和福利,推动经济的稳定和可持续发展。在新的时代背景下,保险行业面临着新的机遇和挑战,需要不断创新和转型,以适应社会需求的变化和经济结构的调整近年来,中国的保险科技蓬勃发展。通过创新的商业模式和先进的技术手段,积极推动保险行业实现数字化和智能化转型。保险科技的目标是提升保险服务的便利性、个性化和智能化水平,以前所未有的速度改变传统保险业的面貌。这一发展趋势为保险行业注入了新的活力,使保险产品更贴近人民群众的实际

快手黑科技“子弹时间”赋能亚运转播,打造智慧观赛新体验快手黑科技“子弹时间”赋能亚运转播,打造智慧观赛新体验Oct 11, 2023 am 11:21 AM

杭州第19届亚运会不仅是国际顶级体育盛会,更是一场精彩绝伦的中国科技盛宴。本届亚运会中,快手StreamLake与杭州电信深度合作,联合打造智慧观赛新体验,在击剑赛事的转播中,全面应用了快手StreamLake六自由度技术,其中“子弹时间”也是首次应用于击剑项目国际顶级赛事。中国电信杭州分公司智能亚运专班组长芮杰表示,依托快手StreamLake自研的4K3D虚拟运镜视频技术和中国电信5G/全光网,通过赛场内部署的4K专业摄像机阵列实时采集的高清竞赛视频,

See all articles

Hot AI Tools

Undresser.AI Undress

Undresser.AI Undress

AI-powered app for creating realistic nude photos

AI Clothes Remover

AI Clothes Remover

Online AI tool for removing clothes from photos.

Undress AI Tool

Undress AI Tool

Undress images for free

Clothoff.io

Clothoff.io

AI clothes remover

AI Hentai Generator

AI Hentai Generator

Generate AI Hentai for free.

Hot Tools

SublimeText3 Mac version

SublimeText3 Mac version

God-level code editing software (SublimeText3)

MantisBT

MantisBT

Mantis is an easy-to-deploy web-based defect tracking tool designed to aid in product defect tracking. It requires PHP, MySQL and a web server. Check out our demo and hosting services.

MinGW - Minimalist GNU for Windows

MinGW - Minimalist GNU for Windows

This project is in the process of being migrated to osdn.net/projects/mingw, you can continue to follow us there. MinGW: A native Windows port of the GNU Compiler Collection (GCC), freely distributable import libraries and header files for building native Windows applications; includes extensions to the MSVC runtime to support C99 functionality. All MinGW software can run on 64-bit Windows platforms.

WebStorm Mac version

WebStorm Mac version

Useful JavaScript development tools

Safe Exam Browser

Safe Exam Browser

Safe Exam Browser is a secure browser environment for taking online exams securely. This software turns any computer into a secure workstation. It controls access to any utility and prevents students from using unauthorized resources.