Sentiment Analysis of Indonesian Audio Transcription Results Using T5 (Text-to-Text Transfer Transformer)

Abstract

In the digital era, sentiment analysis has become a vital tool for understanding public opinion, particularly from data derived from digital media such as videos. However, voice-based sentiment analysis in the Indonesian language remains uncommon. This research aims to develop the T5 model for sentiment analysis of Indonesian generated from speech using speech-to-text technology. The primary advantages of the T5 model lie in its ability to process lengthy texts, comprehend natural language context, and adapt training for specific tasks such as sentiment analysis. The research dataset was obtained from 20 YouTube videos, segmented into clips of a maximum duration of 15 seconds,  resulting in a total of 300 sentences consisting of 150 positive sentiments and 150 negative sentiments. The generated text data was processed using the T5 model, which was specifically trained to detect positive and negative sentiments through the optimization of specific hyperparameters. The results demonstrated that the T5 model achieved an accuracy of 83%, with a precision of 0.85, a recall of 0.83, and an F-measure of 0.83 when tested on datasets different from the training data. This research indicates that the T5 model can be adapted for voice-based sentiment analysis in the Indonesian language with satisfactory results. These findings contribute to the development of voice-based sentiment analysis technology, which can be applied to opinion analysis or product reviews. In the future, improving the pre-processing stage and using more diverse datasets are expected to improve the overall performance of the model.