Speech emotion recognition using speech feature and word embedding

Emotion recognition can be performed automatically from many modalities. This paper presents a categorical speech emotion recognition using speech feature and word embedding. Text features can be combined with speech features to improve emotion recognition accuracy, and both features can be obtained from speech. Here, we use speech segments, by removing silences in an utterance, where the acoustic feature is extracted for speech-based emotion recognition. Word embedding is used as an input feature for text emotion recognition and a combination of both features is proposed for performance improvement purpose. Two unidirectional LSTM layers are used for text and fully connected layers are applied for acoustic emotion recognition. Both networks then are merged by fully connected networks in early fusion way to produce one of four predicted emotion categories. The result shows the combination of speech and text achieve higher accuracy i.e. 75.49% compared to speech only with 58.29% or text only emotion recognition with 68.01%. This result also outperforms the previously proposed methods by others using the same dataset on the same modalities.