ROMJIST Volume 21, No. 4, 2018, pp. 460-474
Eduard FRANȚI, Monica DASCĂLU, Ioan ISPAS, Ana Voichita TEBEANU, Zoltan ELTETO, Silvia BRANEA, Voichita DRAGOMIR Decoding communication: a deep learning approach to voice-based intention detection
ABSTRACT: This paper presents an original method of intention detection that can open a new direction of research in voice-based affective computing. A deep learning approach was used to detect the consistency between real and expressed intentions of a speaker (or the inconsistency, that is related to deceiving – or manipulative – intention), as reflected in the voice. The labelling and triangulation of results implies a qualitative research method, critical discourse analysis, and require expert evaluation. The method was implemented in a software platform integrated with the neural network programming frame. The deep learning architecture selected is based on similar models used by the authors in affective computing applications. The experimental researched applied the proposed method for a famous historical case: US President Richard Nixon’s audio speeches from the ‘Watergate affair’. A labelled data base of 2758 files (2 seconds audio fragments) was generated, based on publicly available voice recordings of President Nixon. These files were used for training and tests and an accuracy of over accuracy of over 94% was obtainedRead full text (pdf)