Technical Name A speech recognition method for detecting fraud
Project Operator National Taipei University of Technology
Project Host 鍾建屏
Summary
API-ready platform combining speech/text emotion analysis, Transformer-based intent extraction and ten-pattern deception scoring to streamline claims and lending. A 3-D PAD dashboard visualises customer affect, while generative sample balancing and continual learning keep fraud detection sharp—delivering faster, safer and more personalised financial services.
Scientific Breakthrough
The invention fuses MFCC-TIM-Net speech affect, GPT-based latent text sentiment, Transformer intent graphs and PAD-Gaussian 3-D emotion surfaces into one multimodal engine. A ten-pattern deception ontology converts dialogues into quantitative trust metrics, bridging computational paralinguistics, NLP and fraud psychology.
Industrial Applicability
Plug-and-play APIs let insurers, banks and SaaS call-centres stream dialogues into the platform, which returns live 3-D affect plots and risk scores. Pilots cut claim/loan cycles from 7 days to 2, trimmed call handling by 12 %, and saved NT$6 M per 100 k interactions. Flexible SaaS or on-prem licensing scales to KYC, collections and public hotlines, with ROI in under eight months.
Keyword Multimodal emotion-semantic analytics TIM-Net speech affect recognition GPT text sentiment Transformer intent extraction PAD-Gaussian 3-D affect visualisation Ten-pattern deception classification Generative data augmentation Insurance & loan risk control Real-time RESTful APIs Edge deployment
  • Contact
  • Chien-Ping Chung