Abstract

Emotion recognition from speech got an important research field over the last years. It was noticed very early that the same methods used for speech recognition can also be used for emotion recognition.
But emotion recognition from speech has still lots of unresolved problems.

1) The recognition rate for not acted emotions decreases when using more than 4-6 emotion classes. To improve the performance some research try to investigate which features perform best. But another approach would be to examine where to look for emotions within an utterance. Speech recognition distinguishes between silence and words/phonemes. Emotion recognition tries the same while applieing a neutral model, that feels responsoble for most of the utterances.
Should we use another mechanism for emotion recognition? How would it look like?
The talk will present some preliminary results for this questions dealing with human-annotated material.

2) As mentioned in other SFB-talks before it is still an open question which emotions are relevant for human-machine-communication and how they could be represented. Todays used basic emotions derived from Ekman or the Geneve Wheel of Emotions or the Self Assesment Manikins?
Therefore this talk will present an comparison of three different emotion theories on humal labelled audio-material from human-machine-communication.