Electronics & Telecommunications student at
Hanoi University of Science and Technology
Digital Signal Processing · Biomedical Signal Processing · TinyML on the Edge
I build signal-processing and machine-learning pipelines for biomedical audio and physiological signals, with a focus on making them run on resource-constrained edge devices.
- Audio source separation — separating overlapping heart and lung sounds
- ECG analysis — arrhythmia detection from single-lead recordings
- Respiratory sound classification — screening for COPD from lung auscultation
- Embedded deployment — ESP32, Jetson, TensorFlow Lite
The 43rd Student Research Conference — HUST (2025–2026)
Artificial Intelligence-Based Audio Source Separation Techniques for Edge Computing Kỹ thuật phân tách âm thanh sử dụng trí tuệ nhân tạo tại biên
| Project | Description | Stack |
|---|---|---|
| SPMamba-HLS | Heart/lung sound separation on HLS-CMDS using SPMamba | PyTorch, Mamba |
| Detect-spelling-errors | Two-stage Vietnamese spell checker — Transformer detector + LLM verifier | Flask, HuggingFace |
| Cinelune-main-prj | Movie streaming web app with personalized recommendations | React, Express, SQLite |
| SpO2-Monitor | Blood oxygen saturation monitoring | Python |
| Goods-management-software | Inventory management desktop application | C# |
- ECG arrhythmia classification (PhysioNet/CinC 2017) — 28-feature LightGBM pipeline with isotonic calibration. Hold-out AUC 0.916, Se 0.851 / Sp 0.837. Real-time deployment on ESP32 + AD8232.
- MIT-BIH beat classification — HRV, morphology and template-correlation features; GroupKFold to prevent patient-level leakage; PyQt5 live monitoring GUI.
- Lung sound classification for COPD — 6-class TFLite model at 94.4% accuracy, targeting Jetson deployment.
- Environmental sound recognition (ESC-10) — LTI filter-bank baseline (37.5%) vs. CNN (81.25%).
Domains: digital signal processing · audio source separation · biomedical signal processing · TinyML & edge deployment · RTL digital design (Verilog/Vivado)