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•学生论坛丨杨东超：面向多任务的音频基座模型：音频生成视角
形式：线上
时间：6月6日(周五) 14:00 ~ 16:00
报告嘉宾
杨东超，香港中文大学在读博士生，导师为Helen Meng教授，他的研究方向为多任务音频基础模型，致力于在统一框架下整合speech、music及audio等多种音频处理任务。他最近工作关注于音频离散化（tokenization）、层次化建模，以及与大型语言模型的跨模态融合。
报告摘要
In this talk, Dongchao Yang will present his recent progress in building multi-task audio foundation models—unified systems that can handle a wide range of audio tasks, from speech-related task to music and sound generation. Inspired by large language models, Yang's team's approach treats audio as a sequence of discrete tokens, enabling the use of transformer-based models for both understanding and generation. He will first introduce their work for building a low bit-rate and semantic-rich audio tokenizer. In this work, they present a query-based compression paradigm for audio codec model. Then he will introduce their hierarchical modeling strategy to effective model long audio sequences in one stage, and show the effectiveness of building a multi-task audio foundation model. Lastly, he will present a cross-modality audio tokenizer that bridges audio and text, allowing few-shot learning across tasks. These contributions mark a step forward toward universal audio intelligence.
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：661-917-700
