Gender-dependent emotion recognition based on HMMs and SPHMMs

It is well known that emotion recognition performance is not ideal. The work of this research is devoted to improving emotion recognition performance by employing a two-stage recognizer that combines and integrates gender recognizer and emotion recognizer into one system. Hidden Markov Models (HMMs) and Suprasegmental Hidden Markov Models (SPHMMs) have been used as classifiers in the two-stage recognizer. This recognizer has been tested on two distinct and separate emotional speech databases. The first database is our collected database and the second one is the Emotional Prosody Speech and Transcripts database. Six basic emotions including the neutral state have been used in each database. Our results show that emotion recognition performance based on the two-stage approach (gender-dependent emotion recognizer) has been significantly improved compared to that based on emotion recognizer without gender information and emotion recognizer with correct gender information by an average of 11% and 5%, respectively. This work shows that the highest emotion identification performance takes place when the classifiers are completely biased towards suprasegmental models and no impact of acoustic models. The results achieved based on the two-stage framework fall within 2.28% of those obtained in subjective assessment by human judges.