Implementasi Algoritma LSTM pada Speech Recognition dalam Aplikasi Mobile untuk Membantu Tunanetra Membaca Dokumen

Abstract

Blind and visually impaired individuals face challenges in accessing printed documents due to the limited availability of braille formats and the high cost of reading aids. This accessibility gap restricts their independence in obtaining information. This study aims to implement a recurrent neural network–based algorithm, specifically Long Short-Term Memory (LSTM), for the speech-recognition feature in a mobile application designed to help visually impaired users read documents. The app integrates Optical Character Recognition (OCR), Text-to-Speech (TTS), and voice commands. The primary focus of the research is the development of the voice-command feature, enabling users to operate the application independently without relying on others. The command-speech dataset used in this research consists of recordings of “Foto” (Photo), “Info” (Info), “Baca” (Read), “Ulang” (Repeat), “Berhenti” (Stop), and “Kembali” (Back), from 53 male and female respondents across various age ranges. The data undergo preprocessing steps—including audio loading, standardization, noise reduction, and band-pass filtering—followed by extraction of Mel-Frequency Cepstral Coefficients (MFCC), label encoding, and padding before being fed into the LSTM model. The best model in this study achieved a testing accuracy of 96.6%. Implementation is carried out using FastAPI to connect the Android mobile application with the speech recognition model. User testing with five visually impaired participants for each test yielded a User Satisfaction Score (USS) of 4.4 out of 5.