|
|
|
Improved Emotion Recognition with Novel Global Utterance-level Features |
|
PP: 147S-153S |
|
Author(s) |
|
Yongming Huang,
Guobao Zhang,
Xiong Li,
Feipeng Da,
|
|
Abstract |
|
Traditional features, which are extracted from each frame, can not reflect the dynamic characteristics of emotion speech signal accurately. To solve this problem, first, without dividing the emotion speech into frames, novel global utterance-level features are proposed with multi-scale optimal wavelet packet decomposition; then, considering the case of little training samples, a fusion strategy through metric learning, which is called weak metric learning in this work, is proposed for fusing the global and traditional features. The experimental results with LIBSVM show that fusing the novel global feature to traditional feature achieves significant improvements about 5.2% to 13.6% than merely using local utterance-level features. |
|
|
|
|