About / 个人简介

I am a 2023-cohort undergraduate in the Radio Wave Propagation and Antennas Honors Program at the School of Physics, Xidian University, and expect to graduate in July 2027. My research lies at the intersection of multimedia signal processing and artificial intelligence, with a focus on multimodal generation and understanding, deepfake speech detection, model generalization, and interpretability.

I have been admitted to the Ph.D. program in the Department of Engineering Physics at Tsinghua University and will be advised by Prof. Junjie Zhang. My previous research experience includes studying the generalization of deepfake speech detectors across speech-generation paradigms in Academician Xiaoping Zhang’s group, leading the VMM research on video-to-music generation as first author, and contributing to HumanMAC++ on diffusion-based human motion prediction.

许嘉俊，西安电子科技大学物理学院电波传播与天线实验班 2023 级本科生，主要研究多模态生成与理解、深伪语音检测，以及模型泛化与可解释性。已保送至清华大学工程物理系攻读博士，师从张俊杰老师。此前在清华大学张晓平院士课题组开展深伪语音检测研究，关注检测模型在不同语音生成范式下的泛化能力；以第一作者身份主导 VMM 视频到音乐生成研究，并参与基于扩散模型的人体运动预测项目 HumanMAC++。

Research Interests

Trustworthy multimedia intelligence（可信多媒体智能）: deepfake speech detection, out-of-distribution generalization, failure analysis, and model interpretability.
Multimodal generation and understanding（多模态生成与理解）: video-to-music generation, audiovisual semantic and affective alignment, and generative models.
Physics-inspired foundations of large models（物理启发的大模型理论基础）: exploring physical principles for energy-efficient LLMs and AGI-oriented learning systems, while using mathematical tools to formulate rigorous, operational definitions of intelligence.

News

October 2026: Received the National Scholarship for Undergraduate Students (Top 1%).
September 2026: Submitted What survives the codec shift: Pooled no-vocals residuals for speech deepfake detection to ICASSP 2027.
May 2026: Received an Honorable Mention in the Mathematical Contest in Modeling (MCM).
March 2026: Joined Xiaoping Zhang's group at Tsinghua University to work on deepfake speech detection.
October 2025: The first-author VMM paper became available online in Computer Vision and Image Understanding.
2024: Received the National Scholarship for Undergraduate Students and the Outstanding Student Model honor at Xidian University.

Publications

Overview of the VMM video-to-music generation framework

VMM: Video-Music Mamba for Generating Background Music from Videos

Jiajun Xu, Zixiang Lu, Ping Gao, Qiguang Miao, Kun Xie
Xidian University
Computer Vision and Image Understanding, 262, Article 104545, December 2025.

Paper Project Code

VMM combines Mamba and Transformer components for video-to-music generation, coordinating video and chord conditioning with a Switch Schedule.

Overview of the MN-P dual-view speech deepfake detection framework

What Survives the Codec Shift: Pooled No-Vocals Residuals for Speech Deepfake Detection

Jiajun Xu, Menglu Li, Xiao-Ping Zhang
Tsinghua University
arXiv preprint, arXiv:2609.33375, September 2026.

Paper

MN-P combines pooled no-vocals residuals with token-level XLS-R features through adaptive gating for codec-unseen speech deepfake detection.

Research

Deepfake Speech Detection

Research Intern, Xiaoping Zhang’s Group, Tsinghua University | March 2026 - September 24, 2026

I studied how deepfake speech detectors generalize across speech-generation paradigms. The work developed controlled evaluation and representation-diagnostic procedures and explored multi-layer speech representations with Mixture of Experts integration for unseen codec conditions.

VMM: Video-Music Mamba

First Author | June 2024 - November 2025

I led the problem formulation, literature review, method design, experiments, visualization, and paper writing for video-to-music generation. The project investigates visual semantics, temporal structure, and affective alignment in generated music.

HumanMAC++

Research Intern, Tsinghua University | September 2023 - December 2023

I contributed to experimental validation for diffusion-based human motion prediction and studied the SDE view that unifies NCSN and DDPM, together with the DDNM method.

Diffusion-Model-Based Mural Restoration Platform

Project Lead, National Undergraduate Innovation and Entrepreneurship Training Program

I led the team through literature study, method design, model reproduction, system implementation, debugging, and final presentation. The project completed its planned review and explored generative models for digital cultural-heritage restoration.

Education

Tsinghua University

Incoming Ph.D. Student, Department of Engineering Physics | Expected Enrollment: 2027

Xidian University

B.Sc. Student, Radio Wave Propagation and Antennas Honors Program | 2023 - Expected July 2027

GPA: 3.84/4.00
Rank: 1/33 (final program ranking)

Honors and Awards

2026 National Scholarship for Undergraduate Students (Top 1%)
2026 Mathematical Contest in Modeling (MCM), Honorable Mention
2024 Interdisciplinary Contest in Modeling (ICM), Honorable Mention
2024 National Scholarship for Undergraduate Students (Top 1%)
2024 Outstanding Student Model, Xidian University

Last updated: October 2026

Jiajun Xu（许嘉俊）