Study on Feature Extraction Method from 2D Character Illustration based on Human’s Cognitive Characteristics for Automatic Voice Estimation

Abstract

Humans can imagine an approximate voice from a human face, and some studies estimate and generate a voice from a human face. As research applying this, there is research to create a sound from a 2D illustration character. This study considers how a person imagines a voice from a face and examines a method for generating speech from a 2D illustration character. So far, we have verified which voice actor/actress resembles that character’s voice from an illustration of an unknown character by learning by associating a character with a voice actor. As a result, 2 out of 5 characters were judged correctly. We also conducted a questionnaire on where people look in the character’s illustrations to imagine their voices and found that their eyes and hair are powerful features. In consideration of the above results, this study attempts to acquire eye information from 2D illustrated characters. The system improved the extraction system by rotating the image by detecting landmarks to extract the character’s eyes. As a result, when the detection accuracy was verified, the result was 73.7