Implementasi Deteksi Bacaan Hijaiyah menggunakan Model Pra-latih Pengenalan Suara Berbasis Contrastive Learning pada Aplikasi Belajar Tilawati

Prasetyo, Muhammad Mirza Ralfie (2026) Implementasi Deteksi Bacaan Hijaiyah menggunakan Model Pra-latih Pengenalan Suara Berbasis Contrastive Learning pada Aplikasi Belajar Tilawati. Other thesis, Institut Teknologi Sepuluh Nopember.

Text 5025221220-Undergraduate_Thesis.pdf - Accepted Version Restricted to Repository staff only Download (10MB) | Request a copy

Abstract

Membaca Al-Qur'an dengan baik dan benar merupakan kewajiban umat Muslim, dan metode Tilawati menjadi salah satu metode pembelajaran yang umum digunakan di Indonesia. Namun, metode ini sangat bergantung pada evaluasi guru sehingga kurang efektif untuk pembelajaran mandiri. Oleh karena itu, diperlukan solusi berupa aplikasi belajar Tilawati sebagai alat evaluasi mandiri untuk berlatih melafalkan huruf hijaiyah dengan baik dan benar. Penelitian ini bertujuan mengembangkan aplikasi Tilawati berbasis mobile untuk mendeteksi bacaan hijaiyah menggunakan Wav2Vec2.0 XLSR yang merupakan model pralatih pengenalan suara berbasis contrastive learning yang dilatih dengan data suara berukuran besar. Pada penelitian ini digunakan dataset rekaman pelafalan huruf hijaiyah dengan harakat fathah pada 26 murid TPQ Zidna Ilma di Depok dan dataset bacaan hijaiyah berharakat fathah open source dari Kaggle. Model Wav2Vec2.0 XLSR dilatih melalui transfer learning dua tahap yaitu pemilihan encoder terbaik melalui eksperimen klasifikasi huruf hijaiyah, dilanjutkan finetuning menjadi model Automatic Speech Recognition (ASR) dengan lapisan Connectionist Temporal Classification (CTC) dan kosakata tingkat kata. Performa model dievaluasi menggunakan metrik Character Error Rate (CER) dan Word Error Rate (WER). Encoder Wav2Vec2.0 Jonatasgrosman (model yang sudah di Fine Tune dalam Bahasa Arab) terpilih sebagai backbone dengan akurasi klasifikasi 93,07% pada data validasi. Model ASR final mencapai CER 3,82%, WER 7,12%, serta akurasi 92,8% pada data uji. Pengujian pada lingkungan nyata oleh sepuluh pembaca menghasilkan rata-rata CER 14,46% yang dipengaruhi variasi jeda antar huruf dan umur pembaca. Model berhasil diintegrasikan ke aplikasi mobile dan lolos seluruh pengujian fungsional.
=====================================================================================================================================
Reading the Qur'an properly and correctly is a religious obligation for Muslims, and the Tilawati method has become one of the commonly used learning methods in Indonesia. However, this method is heavily dependent on teacher evaluation, making it less effective for self-directed learning. Therefore, a solution is needed in the form of a Tilawati learning application that serves as an independent evaluation tool for practicing the correct pronunciation of Hijaiyah letters. This research aims to develop a mobile-based Tilawati application to detect Hijaiyah letter pronunciation using Wav2Vec2.0 XLSR, a pre-trained speech recognition model based on contrastive learning that is trained on large-scale voice data. The study utilizes a dataset of Hijaiyah letter pronunciation recordings with fathah diacritics from 26 students at TPQ Zidna Ilma in Depok, and an open-source Hijaiyah reading dataset with fathah diacritics from Kaggle. The Wav2Vec2.0 XLSR model is trained through two-stage transfer learning: first, selecting the best encoder through Hijaiyah letter classification experiments, followed by fine-tuning into an Automatic Speech Recognition model with Connectionist Temporal Classification (CTC) layers and word-level vocabulary. Model performance is evaluated using Character Error Rate (CER) and Word Error Rate (WER) metrics. The Wav2Vec2.0 Jonatasgrosman encoder was selected as the backbone, achieving 93.07% classification accuracy. The final ASR model achieved 3.82% CER, 7.12% WER, and 92.8% accuracy on the test data. Real-world testing by ten readers resulted in an average CER of 14.46%, influenced by variations in inter-letter pauses and reader age. The model was successfully integrated into the mobile application and passed all functional testing.

Item Type:	Thesis (Other)
Uncontrolled Keywords:	Contrastive Learning, Hijaiyah, Tilawati, Aplikasi Mobile, Pengenalan Ucapan, Wav2Vec2.0, Contrastive Learning, Hijaiyah, Tilawati, Mobile App, Speech Recognition, Wav2Vec2.0
Subjects:	Q Science > QA Mathematics > QA76.585 Cloud computing. Mobile computing. Q Science > QA Mathematics > QA76.758 Software engineering T Technology > TK Electrical engineering. Electronics Nuclear engineering > TK7882.S65 Automatic speech recognition.
Divisions:	Faculty of Intelligent Electrical and Informatics Technology (ELECTICS) > Informatics Engineering > 55201-(S1) Undergraduate Thesis
Depositing User:	Muhammad Mirza Ralfie Prasetyo
Date Deposited:	24 Jul 2026 03:01
Last Modified:	24 Jul 2026 03:01
URI:	http://repository.its.ac.id/id/eprint/136705

Actions (login required)

View Item