Xmart青年论坛由上海交通大学跨媒体语言智能实验室（SJTU X-LANCE lab）创办，中国计算机学会语音对话专委会主办，语音之家协办，旨在邀请国内外优秀的青年学者分享其最新科研工作和成果，促进多元且深入的交流与合作。Xmart学生论坛作为其中一个系列，致力于邀请国内外知名高校有成体系工作的研究生，主要通过线上分享的方式，系统地介绍其科研成果和心得，为青年学生打造一个学术探讨，思维碰撞和多学科交叉融合的平台。
Xmart•学生论坛丨王远程：迈向自然且高效的语音合成
形式：线上
时间：6月26日(周四) 14:00 ~ 16:00
报告嘉宾
王远程，香港中文大学（深圳）在读博士生，师从武执政教授。主要研究方向包括语音生成、语音表征学习以及音频大模型。他在 NeurIPS、ICML、ICLR、ACL 等国际顶级学术会议上发表多篇论文。同时，他是开源音频生成工具包 Amphion 的主要发起人和核心代码贡献者之一，GitHub 累计获得 9.2K+ stars，开源了包括 MaskGCT、NaturalSpeech 3、FACodec 和 AUDIT 在内的多个语音音频合成模型。
报告摘要
Achieving natural and efficient speech synthesis is a key challenge in generative AI and serves as a fundamental building block for realistic human–machine interaction. In this talk, I will present our recent explorations toward building more natural, robust, and efficient speech synthesis systems from three complementary perspectives.
(1) Modeling: I will introduce our exploration of masked language modeling (a.k.a. discrete diffusion models) approaches for speech generation. This includes a brief revisit of our prior work, MaskGCT, as well as our recent effort on developing a foundation model for unified speech synthesis through masked generative pre-training.
(2) Alignment (post-training): I will present our ACL 2025 work, Advancing Zero-shot Text-to-Speech Intelligibility across Diverse Domains via Preference Alignment, where we explore how preference alignment can enhance intelligibility under challenging conditions, such as cross-lingual and code-switching scenarios in speech generation.
(3) Representation: Lastly, from the representation learning perspective, I will introduce TaDiCodec, our latest work on building a discrete speech tokenizer with a single codebook and an extremely low token rate of 6.25 Hz, designed for speech generation and beyond.
参加方式
①
直播将通过语音之家微信视频号进行直播
手机端、PC端可同步观看
👇👇👇
②
腾讯会议参加
会议号：681-969-050
