Implementasi Speech Recognition Menggunakan Long Short-Term Memory untuk Software Presentasi

Abstract

Presentation is one of the methods for delivering thoughts, ideas, and concepts to an audience verbally. Presentation activities can be supported by presentation software that can be used to organize the sequence of material to be presented with visually appealing visuals. Operating presentation software requires technical assistance such as a remote, mouse, keyboard, and even a personal assistant, which can be distracting to the presenter as it limits their freedom in delivering the material. This distraction can be addressed through the implementation of speech recognition as a command to operate presentation software, making it easier for the presenter. A speech recognition system is developed using Long Short-Term Memory (LSTM), which can handle the issues of long-term dependency and vanishing gradient associated with Recurrent Neural Networks (RNN). There are 10 command words used to operate the presentation software. LSTM demonstrates superior performance when compared to alternative techniques like DNN, CNN, and SimpleRNN, achieving a training accuracy of 96.5%, a validation accuracy of 94.8%, and a testing accuracy of 94%. The LSTM method can be effectively used for sequential data to recognize real-time speech.