An AI-Based Lie Detector for Call Center Conversations
upstart writes:
An AI-Based Lie Detector for Call Center Conversations:
Researchers in Germany have used machine learning to create an audio analysis system intended primarily to act as an AI-based lie detector for customers in audio communications with call center and support staff.
The system uses a specially-created dataset of audio recordings by 40 students and teachers during debates on contentious subjects, including the morality of the death penalty and tuition fees. The model was trained on an architecture that uses Convolutional Neural Networks (CNNs) and Long Short-Term Memory (LSTM), and achieved a reported accuracy rate of 98%.
Though the stated intent of the work cites customer communications, the researchers concede that it effectively operates as a general purpose lie-detector:
[...] In the absence of a suitable publicly available dataset in the German language, the researchers - from Neu-Ulm University of Applied Sciences (HNU) - created their own source material. Fliers were posted at the university and at local schools, with 40 volunteers selected with a minimum age of 16. Volunteers were paid with a 10 euro Amazon voucher.
The sessions were conducted on a debate club model designed to polarize opinion and arouse strong responses around incendiary topics, effectively modeling the stress that can occur in problematic customer conversations on the phone.
Journal Reference:
Fabian Thaler, Stefan Fauer, Heiko Gewald. Put your money where your mouth is: Using AI voice analysis to detect whether spoken arguments reflect the speaker's true convictions, (DOI: https://arxiv.org/abs/2107.11175)[pdf]
Read more of this story at SoylentNews.