EMOTION RECOGNITION MODEL FROM INDONESIA SPOKEN LANGUAGE USING ACOUSTIC AND LEXICAL FEATURE

• Main Author: KASYIDI (NIM : 23515025), FATAN
• Format: Theses
• Language: Indonesia
• Online Access: https://digilib.itb.ac.id/gdl/view/22088
• Institution: Institut Teknologi Bandung

Emotion recognition from utterance signal have been developed in many languages include Indonesia spoken language. There are four class emotion that can be recognized by a system in Indonesia spoken language such as angry, happiness, contentment and sadness. These are considered as general emotion that use in daily activities. <br /> <br /> This thesis developed a model for emotion recognition from utterance signal in Indonesia spoken language using acoustic and lexical features. Acoustic features can be mention such as spectral, ceptral, voicing-related, energy, pitch contour, jitter and shimmer. Beside that, lexical feature that use in this thesis such as Bag-of-word and TF-IDF. From that kind of feature, there will be three groups of model that consist of acoustic model, lexical model and combined model from both of features. These features were extracted from emotion corpus that developed alongside the transcript from the audio file. <br /> <br /> There are 100 segmen of utterance as data testing with four class emotion. Data development are use to develop the model of three with 982 segmen of utterance in Indonesia spoken language. The experimental scenario for data development is using 5-fold cross validation and three machine learning algorithm such as naive bayes, random forest and support vector machine (SVM) will included. The result of experiment shows that acoustic model have F-measure 0.40361 by using random forest, F-measure 0.40823 by using random forest and combined model have F-measure 0.40823 by using SVM RBF kernel. Finally the result from testing phase by use the data testing show that the best accuracy given by combined model between acoustic and lexical TFIDF using SVM RBF kernel with 0.476 of F-measure.