陆奕衡 Yiheng Lu

I contributed to advancing filler-word recognition and prosody-aware speech processing through feature engineering, dataset evaluation, and data preparation for synthesis.

I proposed a timestamp-based filler and prosody feature extraction method leveraging WhisperX, FunAudio, and Whisper-timestamped to improve recognition accuracy.
I further conducted cross-dataset evaluation on Emilia and EARS MM_TTS, validating robustness across English long-form audio and daily conversations.
Additionally, I generated over 100 text slices per filler label to support TTS synthesis and evaluation, and designed the NVSpeech project homepage to present the system architecture, demos, and datasets.

I led my team to leverage the latest AI technologies to build a system that helps novice academic researchers with paper finding, summarization, and retrieval.

We critically analyzed the limitations of current academic search engines and conventional RAG systems.
I proposed an improved architecture and coordinated the division of tasks among team members.
My role focused on backend development, including model fine-tuning, and the integration of MCP, large language models (LLMs), and database systems to ensure seamless functionality.

I led my team to present a hybrid method for segmenting lung lesions in COVID-19 CT scans. It combines classical image processing with deep learning to detect 3 common lesion types.

HyCoSeg is a hybrid pipeline that integrates classical image processing with deep learning for the segmentation of COVID-19 lung lesions in CT scans. The lung regions are first delineated with a 2D U-Net (LungMask) to constrain subsequent detection. Ground-glass opacities are segmented through Hounsfield Unit thresholding combined with morphological post-processing, while consolidation is identified using a coarse-to-fine deep model enhanced by Gabor texture filtering. Pleural effusion is extracted by boundary-aware density thresholding followed by connected-component analysis.