Development of a Speech Emotion Recognition Dataset for Indonesian

Speech emotion recognition (SER) is a crucial technology for applications such as human-computer interaction, mental health monitoring, and call centre analytics. However, SER research in Indonesia is still limited by a lack of annotated speech emotion datasets. In this study, we provided EEPIS’s Speech Emotion Recognition Audio Video Dataset (E-SERAVD), an Indonesian speech emotion recognition dataset that has four basic emotions: Happy, Angry, Neutral, and Sad. The data obtained from Indonesian movies and scenes were trimmed to capture emotional expressions. The dataset creation process is based on a model widely used by previous researchers. In this case, dataset creation begins with the process of defining the parameters and types of audio recording to be targeted. The next step is data collection, followed by the data cleaning process. The data validation stage involves a testing process by three raters to prepare the labelling process. The three raters’ observations showed satisfactory performance, with agreement scores of 88.8, 68.8, and 91.4%. The constructed dataset was evaluated using the inter-rater reliability metric, which showed a Fleiss Kappa value of 0.61; therefore, it was included in the good category and was then determined. The final step is to separate the data for training and testing. The baseline model accuracy was evaluated with a value of 90% and an F1 score of 86%. The E-SERAVD Dataset for Indonesian SER was prepared as an open-source dataset, with complete documentation and model benchmarks.