Noisy speech training in MFCC-based speech recognition with noise suppression toward robot assisted autism therapy

Abstract

Autism Spectrum Disorder (ASD) is a group of neurodevelopmental disorders. Autistic children experience challenges in three important areas: social communication, social interaction and repetitive behavior. Robots have become tools to support therapists in autism therapy. Toward integrating social interaction and communication in robot-assisted autism therapy, the robot should have speech recognition ability that can be used in noisy environment. This study investigated the potential of using noisy speech training in MFCC-based speech recognition system with noise suppression toward robot-assisted autism therapy. Experimental results with clean speech training on Japanese speech database suggested that MFCC with noise suppression technique could provide the improvement with significantly higher recognition accuracy than MFCC. With noisy speech training with noise type found in the environment, the performance can be improved even more.