Evie, Li Xinlu

Master's Thesis · 2024-2026

Real-time French medical speech recognition system designed for drone deployment in battlefield environments. Achieves state-of-the-art performance in extreme noise conditions.

📊 Key Results

• Built 50,000-sample dataset with 82.8% real drone noise
• Achieved 7.0% WER (21.7% improvement over baseline)
• 14 out of 21 SNR levels reached 100% accuracy
• Deployed on Raspberry Pi 4B with <3s inference time
• 23x improvement in keyword recognition

💡 Key Finding

Training on drone noise improved performance on ALL noise types:

• Helicopter noise: +7.05% improvement
• Weapon noise: +5.34% improvement
• Radio noise: +3.97% improvement

Geospatial analysis of French medical resources using public health data. Created interactive maps showing regional disparities in healthcare access.

Fine-tuned German BERT on 10,000+ patient reviews for sentiment analysis. Achieved 85% accuracy and analyzed COVID-19 impact on patient satisfaction.