Description

Abstract:

Voice expression recognition is an active research area in recent times. We concentrate on machine learning algorithm to predict the expression of a particular voice signals. Voice signal features and its statistical computations are used to enhance the performance of classifier algorithm. we tackle the problem of affective dimension level recognition by extending the methodology presented in our AVEC 2011 audio subchallenge workshop paper which we showed to outperform the Latent-Dynamic Conditional Random Fields method. In this methodology, the temporal relationships between consecutive levels of a given affective dimension are analysed and modeled by using a Markov model based approach. The affective dimension level recognition problem is solved through a multistage automatic pattern recognition system where the temporal relationships are modeled through the Hidden Markov Model (HMM) framework. Here, we refine the multi-stage classification architecture and test its performance over four affective dimensions (arousal, valence, expectation, dominance) not only on the audio but also on the video data of the AVEC 2011 dataset. We further test it on the pain intensity dimension by using the PAINFUL dataset in order to get a better understanding of the pros and cons of the approach. In particular, we expand the analysis to understand how the duration of affective expressions affects its performance.