Feb 2025 – Jun 2025 · 4 mos

• Focusing on Multimodal Fusion & LLM-enhanced ASR techniques
• 1. Build most downloaded Taiwan open source model: Breeze-ASR-25(https://huggingface.co/MediaTek-Research/Breeze-ASR-25)
• 2. Proposed self-refining framework for ASR training with TTS models, paper accepted by ASRU 2025

Jan 2025 – May 2025 · 4 mos

• Teaching Assistant of “Deep learning and its application”

Sep 2023 – Dec 2024 · 1 yr 3 mos

• Research assistant of productivity optimization lab,
• research about data quality, predictability analysis, and explainable AI.
• Built a large data science pipeline, which contains
• 1. Data preprocessing with SOTA imputatition algorithm, data quality and property assessment
• 2. Active learning interface to re-label and active prompt adjust the model weight.
• 3. Deep learning model with knowledge distillation to reduce 25% model parameters with only 2% performance compensation.
• 4. SHAP model feature explainable module and counterfactual explanation for robust model adjustment
• Working on entropy bound estimation with nn and kernel density estimation methods.

Sep 2024 – May 2025 · 8 mos

• Research Member of NTU EE SPML Lab supervised by Prof. Hung-Yi Lee, focusing on the open-source Taiwan Native Whisper and addressing streaming hallucination issues in ASR models.

May 2024 – Jan 2025 · 8 mos

• Research Member of NTU CSIE IR Lab, Prof. Pu-Jen Cheng, working on research of cross-modality GEN-IR and cold start group recommendation systems.

Feb 2024 – Jul 2024 · 5 mos

• Research member of NTU CSIE Financial Algorithm Lab, Prof. Yuh-Dauh Lyuu.

Jul 2024 – Jun 2025 · 11 mos

• Project support & course design about software development, data engineering, cloud resource management, statistical analytics and deep learning

May 2024 – Sep 2024 · 4 mos

• I process parallel audio data from different devices in Taiwanese and Hakka dialects. Due to the varying start and end times of recordings across devices, and with data from 8 recording devices, it's necessary to uniformly shift and remove silent segments from the beginning or end to avoid affecting the information content.
• The traditional approach is to lock onto one device's data and use Dynamic Time Warping (DTW) as a standard, aligning other devices to it.
• This method consumes enormous space for longer durations. I employed a correlation calculation method, reducing the space complexity to O(n), and used FFT to calculate correlation, achieving a time complexity of O(n log n).
• Co-authored: Channel-aware Domain-Adaptive Generative Adversarial Network for Robust Speech Recognition, submitted to ICASSP 2025
• Using GAN and channel encoder for data augmentation to address the domain mismatch problem in speech recognition.
• This research improves on the original framework for data augmentation targeting out-of-domain noise, used for training denoising systems. Here, an additional encoder is designed for the GAN to incorporate device information. The augmented data then needs to be validated through training on the ASR model. I was responsible for building the ASR model, as well as establishing baselines and toplines on various devices to verify model performance. I also adapted code from different frameworks, porting code from speechbrain and huggingface, to ensure consistency in the experimental setup.

Apr 2024 – Present · 2 yrs 2 mos

• Produce predictive signals (alphas) to employ financial strategies focused on market inefficiencies.

Apr 2024 – Jun 2024 · 2 mos

• Training about the Google Cloud Platform(GCP) Professional Cloud Architect

Jul 2022 – Oct 2022 · 3 mos · On-site

• 1. Improve the log delivery update system with ELK (Elasticsearch - Logstash - Kibana),
• Standardize and automate important information previously stored in respective databases and special mailboxes.
• 2. Develop text parser to automatically capture and parse information from websites, and update it into the format required by the financial analysis group.
• 3. Adjust the large files developed by the company using the waterfall development method and optimize large crawlers → Parser → Database pipeline, updates the data and diversion mode from the early stage, including dynamic web crawlers, Regular Expression, SQL, mongodb and tortoisegit