Machine learning research portfolio
I build trustworthy machine learning systems for audio and music.
Ph.D. Candidate at XJTLU · Expected Mar 2027 · Suzhou, China · Open to 2027 research scientist and machine-learning engineer roles
00 · Focus
My work spans audio understanding, generative music, privacy auditing, and the path from a research method to a creator-facing system.
I study music information retrieval, acoustic scene classification, and perceptually aligned audio evaluation. My work connects signal-level evidence with robust experimental methodology.
I work with diffusion and symbolic-music generation models, including MusicXML-to-audio expressive rendering. The focus is controllable generation and evaluation rather than generation alone.
I develop membership-inference, adversarial, and voice-protection methods for generative audio. Membership inference asks whether a particular work was used to train a model.
As Principal Investigator of a Suzhou concept-verification project, I own model development, macOS product requirements, evaluation design, and user trials for a creator-facing music tool.
01 · Experience
The through-line is methodological: define the claim, design how it can fail, and carry the strongest result toward a usable system.
Develop membership-inference and auditing methods for generative audio and symbolic-music models, with particular attention to reliable identification at low false-positive rates.
Propose generative-inpainting attacks and perceptually aligned evaluation methods that test both technical effectiveness and what listeners can actually hear.
Lead a PoC of the Suzhou Embodied Intelligence Future Education Concept Verification Center, owning expressive-rendering research, macOS product requirements, evaluation design, and user trials.
The system turns one MusicXML score into multiple performance styles and reached TRL 5 through small-scale validation in a relevant setting.
Explore the interactive PoC →Worked on singing-voice generation and enhancement for online karaoke applications through audio simulation, algorithm testing, and evaluation.
02 · Education
Training across connected systems, cyber security, and intelligent engineering underpins my current work in trustworthy audio AI.
B.Eng. · 2017–2021
Internet of Things Engineering
M.Sc. · 2021–2022
Cyber Security Engineering
Ph.D. · 2023–2027 expected
School of Intelligent Engineering
03 · Research
Each project pairs a technical mechanism with an evaluation question that makes the contribution falsifiable.
Membership inference determines whether a work was used in model training. I propose a generative-manifold perturbation method and evaluate identification where false positives must remain low.
Explore the interactive method →MAIA locates critical audio segments and uses generative inpainting to construct white-box and black-box attacks across music information retrieval tasks.
Listen to the examples →A psychoacoustically conditioned representation model evaluates auditory similarity in music adversarial attacks.
Explore the perceptual model →TS-RaMIA turns token-level structural sensitivity into membership evidence, providing an auditable attack pipeline for symbolic music generation models.
Explore the interactive attack →Curriculum-learning methods use uncertainty and multiple training signals to prioritize useful samples when acoustic conditions and data availability change.
Explore the dynamic curriculum →04 · Publications
Selected venues include ICASSP, INTERSPEECH, ISMIR, AAAI Workshop, CMMR, FG, and DCASE.
IEEE International Conference on Acoustics, Speech and Signal Processing
International Society for Music Information Retrieval Conference
AAAI Conference on Artificial Intelligence Workshop · PMLR 303:1–15
International Symposium on Computer Music Multidisciplinary Research
Annual Conference of the International Speech Communication Association
Annual Conference of the International Speech Communication Association
IEEE International Conference on Acoustics, Speech and Signal Processing
IEEE International Conference on Acoustics, Speech and Signal Processing
IEEE International Conference on Automatic Face and Gesture Recognition
Workshop on Detection and Classification of Acoustic Scenes and Events
AAAI Conference on Artificial Intelligence Workshop
Manuscript and technical illustrations · Listed separately from the publication count.
05 · Demos
Ten active research and project pages expose the question, method, evidence, and evaluation boundary; audio examples are included where the underlying work supports them.
Current project page
Locate influential musical regions and regenerate only those bounded segments.
Open verified audio demo →
Current project page
Audit a music diffusion model through local reverse-process stability.
Open evidence-first audit →
Current project page
Use structural-token evidence while controlling evaluation confounders.
Open controlled evidence →
Technical design · Manuscript
Protect live speech before transmission; test voice reuse after multi-clip collection.
Explore the streaming protection design →
Current project page
Turn one MusicXML score into controllable expressive performances.
Open creator-facing PoC →
Current project page
Change sample priorities as learning moves from invariance to potential.
Inspect the dynamic curriculum →
Current project page
Begin with examples that reveal the acoustic scene rather than the recorder.
Inspect the staged curriculum →
Current project page
Represent heart-sound structure across physiological time scales.
Inspect the topological method →
Current project page
Protect public speech with audible background sound that fits the scene.
Inspect the protection pipeline →
Current project page
06 · Toolkit
The toolchain supports reproducible experiments, audio evaluation, research prototypes, and product-oriented iteration.
Diffusion models, music and speech generation, symbolic music modeling, MusicXML workflows
Membership inference, model privacy, copyright auditing, provenance, adversarial robustness
Music information retrieval, audio signal processing, acoustic scene classification, perceptual evaluation
Python, PyTorch, C/C++, MATLAB, experimental design, statistical analysis
Codex, Cursor, and Claude Code for implementation, debugging, prototype iteration, and document automation
Academic writing, peer-reviewed publication, user-trial design, technical presentation
07 · Background
Machine-learning scientist and engineer roles in audio understanding, music intelligence, generative audio, and trustworthy AI.
He expects to complete his Ph.D. in March 2027 and is targeting 2027 graduate or early-career hiring timelines.
Methods for auditing generative audio models, attacking and evaluating music systems, and aligning technical metrics with auditory perception.
Yes. He is Principal Investigator for a MusicXML-to-audio PoC that reached TRL 5 and includes model, app, evaluation, and user-trial ownership.
Yes. At Changba he worked on singing-voice generation and enhancement through audio simulation, algorithm testing, and evaluation.
The selected list appears on this page and the complete record is linked through Google Scholar.
08 · Contact
Available from March 2027 for research scientist and machine-learning engineer roles in Singapore and internationally.