- Korea developed an AI-based voice data analysis model, which has world’s first capability of clustering criminal organizations by tracking criminal’s voice patterns. 
- Compared to existing similar voice analysis models, the accuracy rate has been improved by over 77%.
- Korea plans to spread the model globally through training programs for overseas countries and international conferences.

In response to increasing damage from voice phishing scams globally, Korea developed the world’s first voice data analysis model with the capability to accurately identify criminal voices and cluster criminal organizations. 

The Ministry of the Interior and Safety (MOIS), which developed the model, plans to actively utilize the AI-based voice data analysis model focusing on apprehending voice phishing criminals throughout the entire process of voice crime investigations. 

To prevent voice phishing scams and arrest criminals, the National Forensic Service (NFS) has conducted voice analysis necessary for voice phishing investigations. 

However, there were limitations to the model that NFS was used. In particular, considering that voice phishing criminal organizations often divide roles (investigators, prosecutors, etc.) among group members to deceive victims, it was essential to have the algorithm to cluster criminals involved in voice phishing.

Accordingly, the Integrated Data Analysis Center (IDAC) under MOIS, in collaboration with the NFS, has developed a model capable of grouping those involved in scams with enhanced accuracy of speaker identification since last year, focusing on apprehending voice phishing scammers. 

This model, developed by using the latest AI deep learning technology, learned how to differentiate voice effectively by utilizing approximately one million voice data samples extracted from around 6,000 domestic and foreign speakers.
As follow-up to the development, the model underwent a verification process using over 12,000 voice data samples including actual voices of voice phishing criminals under various scenarios. 

The performance evaluation of the developed analysis model showed a significant improvement in the accuracy of identifying criminals’ voices by about 77%* compared to that of the existing foreign models. 
* (e.g.) Upon analyzing 100 criminal voices, the previous model was only able to discern the similarity of approximately 28 voices. The new model is capable of identifying up to 51 voices. 

The performance evaluation of the developed analysis model showed a significant improvement in the accuracy of identifying criminals’ voices by about 77%* compared to that of the existing other models.  

Using the newly developed models, the IDAC under MOIS and the NFS analyzed 12,323 actual reported criminal voice files and was able to identify the types of repeated crimes, the size of criminal organizations, and the number of crimes commited by each criminal organization.

If you want to know detailed analysis and Korea's plans for utilizing the model, please check the attachment.

Source: Integrated Data Analysis Center