SenseTime Research Institute

ASR & Audio Data Processing Intern · February–May 2025
Supervisor: Shiyin Kang

• GPU Inference Optimization: Migrated a CPU-based remote ASR processing workflow to local GPU inference, optimizing data loading, batching, and processing modules to improve processing efficiency by 50%.
• High-Quality Dialogue Data Selection: Developed an end-to-end data curation pipeline using VAD, ASR, and speech quality metrics (SNR, OVRL, and pause duration) to identify and extract high-quality conversations from large-scale dialogue datasets.
• Named Entity Recognition: Trained a BERT-based NER model to extract artist and song names from music filenames, combining embedding-based similarity matching for duplicate detection and structured metadata construction.