Emotion recognition and speaker identification from speech

The performance of spoken dialogue systems (SDS) is not perfect yet, especially for some languages. Emotion recognition from speech (ER) is a technique which can improve the SDS behavior by finding critical points in the human-machine interaction and changing a dialogue strategy. Inclusion of the speaker specific information, by conducting the speaker identification procedure (SI) at the set up of ER task could also be used in order to improve the dialogue quality. Choosing of both appropriate speech signal features and machine learning algorithms for the ER and SI remain a complex and challenging problem. More than 50 machine learning algorithms were applied in the study for ER and SI tasks, using 9 multi-language corpora (Russian, English, German, and Japanese) of both acted and non-acted emotional utterance recordings. The study provides the results of evaluation as well as their analysis and future directions.