Implementasi MobileNetV2 dan Ekstraksi MFCC dalam klasifikasi pelafalan huruf hijaiyah

Mohammad Putra, Fauzan Fatah (2026) Implementasi MobileNetV2 dan Ekstraksi MFCC dalam klasifikasi pelafalan huruf hijaiyah. Sarjana thesis, UIN Sunan Gunung Djati Bandung.

Preview	Text Cover-1.pdf Download (180kB) | Preview
Preview	Text Abstrak-2.pdf Download (268kB) | Preview
Preview	Text Surat Keterangan bebas plagiarism.pdf Download (379kB) | Preview
Preview	Text Daftar Isi-3.pdf Download (423kB) | Preview
Preview	Text Bab 1-4.pdf Download (533kB) | Preview
	Text Bab 2 - 5.pdf Restricted to Registered users only Download (856kB) | Request a copy
	Text Bab 3 - 6.pdf Restricted to Registered users only Download (722kB) | Request a copy
	Text Bab 4 - 7.pdf Restricted to Registered users only Download (3MB) | Request a copy
	Text Bab 5 - 8.pdf Restricted to Registered users only Download (251kB) | Request a copy
	Text Daftar Pustaka - 9.pdf Restricted to Registered users only Download (260kB) | Request a copy
	Text Lampiran-10.pdf Restricted to Repository staff only Download (4MB) | Request a copy

Abstract

INDONESIA: Rendahnya kemampuan membaca Al-Qur'an di Indonesia 72,25% dari 3.111 responden berada pada level cukup dan kurang menunjukkan kebutuhan akan sistem evaluasi pelafalan yang mampu memberikan umpan balik objektif terhadap pelafalan huruf hijaiyah. Penelitian ini mengimplementasikan arsitektur MobileNetV2 dengan ekstraksi fitur Mel-Frequency Cepstral Coefficients (MFCC) untuk mengklasifikasikan pelafalan huruf hijaiyah, serta membandingkannya dengan Log-Mel Spectrogram sebagai representasi fitur alternatif. Penelitian disusun mengikuti kerangka kerja CRISP-DM dan menggunakan dataset publik berisi 8.904 rekaman audio WAV dari santri Pondok Tahfidz Yanbu'ul Qur'an Menawan, Kudus, yang mencakup 84 kelas (28 huruf hijaiyah × 3 harakat) dengan distribusi seimbang sempurna (106 sampel per kelas). Tahap pra-pemrosesan meliputi noise reduction berbasis spectral gating, trim silence, dan normalisasi durasi menjadi 2 detik. Fitur MFCC (n_mfcc = 40) dibentuk menjadi tiga kanal (statis, delta, delta-delta), sedangkan Log-Mel Spectrogram (n_mels = 128) diduplikasi menjadi pseudo-RGB; keduanya di-resize ke 224×224×3 agar kompatibel dengan input MobileNetV2 pra-terlatih ImageNet. Pelatihan dilakukan dengan augmentasi Specaugment dan strategi transfer learning dua fase pada dua skenario pembagian data (80:10:10 dan 70:15:15). Hasil pengujian menunjukkan model baseline MobileNetV2 + MFCC mencapai test accuracy 77,54%, sedangkan MobileNetV2 + Log-Mel Spectrogram mencapai 86,53% dengan macro precision 87,25%, macro recall 86,47%, dan macro F1-score 86,37%, sekaligus melampaui penelitian acuan (83,80%). Model terbaik diintegrasikan ke dalam aplikasi mobile berbasis FastAPI dan Flutter untuk pengujian pelafalan secara real-time. ENGLISH: The low level of Qur'anic reading proficiency in Indonesia where 72.25% of 3,111 respondents scored at the fair or poor level highlights the need for learning media capable of providing objective feedback on hijaiyah letter pronunciation. This study implements the MobileNetV2 architecture with Mel-Frequency Cepstral Coefficients (MFCC) feature extraction to classify hijaiyah letter pronunciation, and compares it against the Log-Mel Spectrogram as an alternative feature representation. The research follows the CRISP-DM framework and uses a public dataset of 8,904 WAV audio recordings collected from students of Pondok Tahfidz Yanbu'ul Qur'an Menawan, Kudus, covering 84 classes (28 hijaiyah letters × 3 diacritics) with a perfectly balanced distribution of 106 samples per class. Preprocessing consists of spectral-gating noise reduction, silence trimming, and duration normalization to 2 seconds. MFCC features (n_mfcc = 40) are arranged into three channels (static, delta, and delta-delta), while the Log-Mel Spectrogram (n_mels = 128) is stacked into a pseudo-RGB representation; both are resized to 224×224×3 to match the input requirements of ImageNet-pretrained MobileNetV2. Training employs Specaugment augmentation and a two-phase transfer learning strategy across two data-splitting scenarios (80:10:10 and 70:15:15). Experimental results show that the MobileNetV2 + MFCC baseline achieves a test accuracy of 77.54%, whereas MobileNetV2 + Log-Mel Spectrogram reaches 86.53% with a macro precision of 87.25%, macro recall of 86.47%, and macro F1-score of 86.37%, outperforming the reference study (83.80%). The best model is deployed into a FastAPI- and Flutter-based mobile application for real-time pronunciation testing.

Actions (login required)

View Item