抄録
This paper investigates the characteristics of recognizing emotions contained in human speech. The concept of artificial neural networks (ANNs) is adopted for a recognition algorithm. An approach based on the Hidden Markov model (HMM) is also investigated as an alternative recognition method. Using a large database of phoneme-balanced Japanese words, both recognition systems are trained and tested. To evaluate the emotion recognition results obtained by using ANNs or HMMs, emotion recognition testing is carried out with human subjects. The obtained average emotion recognition rates are 51% using ANNs, 32% using HMMs, and 55% with humans. Experimental results confirm that the emotion recognition rate achieved by using the ANNs in the speaker- and context-independent mode is feasible and that ANNs are well suited to this task.
