Deteksi Ucapan Untuk Sistem Pengawasan Asesmen (iProctor) Menggunakan Metode Deep Learning

Juius, Julius (2023) Deteksi Ucapan Untuk Sistem Pengawasan Asesmen (iProctor) Menggunakan Metode Deep Learning. Other thesis, Institut Teknologi Sepuluh Nopember.

Text 05111840000082-Undergraduate_Thesis.pdf - Accepted Version Restricted to Repository staff only until 1 April 2025. Download (2MB) | Request a copy

Abstract

Asesmen adalah kegiatan mengumpulkan informasi tentang ketercapaian kompetensi (rangkaian kemampuan) siswa. Asesmen merupakan bagian integral dari proses pembelajaran, tak terkecuali dalam pembelajaran berbasis Out Door Learning (ODL) dan Massive Open Online Course (MOOC). Perilaku kecurangan dalam pelaksanaan asesmen adalah hal yang tidak dapat dihindari. Sebuah studi menyatakan bahwa persentase siswa yang melakukan kecurangan dalam pelaksanaan kegiatan akademik terus meningkat, dan lebih mudah bagi mereka untuk berlaku curang pada asesmen yang dilakukan secara daring. Hal ini menjadi tantangan untuk perkembangan iProctor, yaitu platform untuk melakukan asesmen seara daring. Untuk mengurangi risiko kecurangan, sistem pelaksanaan dan pengawasan ujian yang valid menjadi suatu hal yang penting. Pada penelitian ini diuji sistem pengawasan otomatis bedasarkan audio. Data audio didapatkan dari mikrofon yang terletak pada ruang dilakukannya asesmen. Sistem pengawasan asesmen dilakukan secara otomatis dengan metode deteksi ucapan menggunakan metode deep learning dengan model CNN. Data audio di ekstraksi fitur menggunakan log mel spectrogram. Hasil esktraksi fitur menjadi input model CNN MobileNetV3. Hasil prediksi dari MobileNetV3 dilakukan proses smoothing dengan metode Majority Vote. Hasil penelitian ini menunjukkan bahwa model deteksi ucapan memberikan hasil terbaik dengan model CNN MobileNetV3-Large pada dataset librispeech dengan speech f1 score 0.8652, non-speech f1 score 0.7332, dan hasil weighted average 0.8242. Ekstraksi fitur menggunakan metode log-mel spectrogram menggunakan parameter fft size 512, mel bins 40, hope size 8, lower frequency 300, upper frequency 8000. Hasil dari log-mel spectrogram dibagi menjadi banyak frame 25ms dan step 12.5ms atau overlap 50.
===============================================================================================================================
Assessment is an activity to obtain information about the achievement of student competencies (a series of abilities). Assessments are an integral part of the learning process, including Out Door Learning (ODL) and Massive Open Online Course (MOOC) based learning. Fraudulent behavior in the implementation of the assessment is possible and unavoidable. A study stated that the percentage of students who cheat in the implementation of academic activities continues to increase, and it is easier for them to cheat in online assessments. This is a challenge for the development of iProctor, which is a platform for conducting online assessments. To reduce the risk of cheating, a valid examination administration and monitoring system is essential. In this study an automatic monitoring system based on audio is tested. Audio data is obtained from a microphone located in the room where the assessment is carried out. The assessment monitoring system is carried out automatically with the speech detection method using the deep learning method with a CNN model. Audio data is feature extracted using the log-mel spectrogram method. The result of feature extraction becomes the CNN MobileNetV3 input model. The prediction results from MobileNetV3 are smoothed. The results of this study indicate that the speech detection model gives the best results with the CNN MobileNetV3-Large model with a speech f1 score of 0.8652, non-speech f1 score of 0.7332, and a weighted average result of 0.8242. Feature extraction uses the log-mel spectrogram method using parameters fft size 512, mel bins 40, hope size 8, lower frequency 300, upper frequency 8000. The results of the log-mel spectrogram are divided into many frames of 25ms and steps of 12.5ms or overlap of 50.

Item Type:	Thesis (Other)
Additional Information:	RSIf 006.454 Jul d-1 2023
Uncontrolled Keywords:	Deep Learning, Deteksi Ucapan, iProctor, Log-mel spectrogram, MobileNetV3
Subjects:	Q Science > QA Mathematics > QA76.6 Computer programming.
Divisions:	Faculty of Intelligent Electrical and Informatics Technology (ELECTICS) > Informatics Engineering > 55201-(S1) Undergraduate Thesis
Depositing User:	Julius Julius
Date Deposited:	08 Feb 2023 07:00
Last Modified:	28 Aug 2023 03:16
URI:	http://repository.its.ac.id/id/eprint/96450

Actions (login required)

View Item