Identifikasi Ucapan Disartria Menggunakan Arsitektur Convolutional Neural Network MobileNetV2 dan MFCC

2025-07-28T10:52:21+00:00

Seiring meningkatnya kebutuhan Automatic Speech Recognition (ASR) pada edge device, pengembangan sistem deteksi gangguan bicara menjadi semakin relevan, terutama untuk aplikasi pada alat kesehatan. Deteksi dini disartria memegang peranan krusial dalam intervensi klinis, sehingga diperlukan model klasifikasi yang efisien agar kedepanya dapat diimplementasikan pada edge atau embedded system untuk membantu proses diagnosis. Penelitian ini bertujuan untuk mengimplementasikan dan mengevaluasi sebuah model deep learning yaitu lightweight Convolutional Neural Network (LCNN) untuk klasifikasi ucapan disartria dengan memanfaatkan arsitektur MobileNetV2 melalui pendekatan transfer learning. Metode penelitian menggunakan dataset publik UASpeech, di mana fitur akustik diekstraksi menggunakan Mel-Frequency Cepstral Coefficients (MFCC) untuk menghasilkan 40 koefisien. Fitur MFCC kemudian divisualisasikan sebagai spektrogram untuk melatih model MobileNetV2 yang sebelumnya telah dilatih pada dataset berskala besar. Hasil evaluasi terbaik pada data uji menunjukkan performa yang sangat baik dengan pencapaian akurasi sebesar 95%

Sistem Rekomendasi Menu Kantin Menggunakan Lifespan-Aware Association Rule Mining Dengan Hybrid Apriori Dan FP-Growth

2025-12-22T05:40:54+00:00

This study addresses the problem of how to systematically increase cross-selling in a small canteen, where additional items such as drinks and snacks are usually offered only based on the cashier’s memory and intuition. The proposed solution is a point-of-sale (POS) recommendation system that suggests complementary menu items in real time, based on patterns learned from historical transaction data. The system uses a lifespan-aware association rule mining approach with a hybrid of Apriori and FP-Growth, where both algorithms are applied to one-hot encoded POS data and their outputs are combined and validated before being deployed as recommendation rules. The research objectives are to extract stable co-purchase patterns from canteen transactions, compare the computational performance of Apriori and FP-Growth in this real-world setting, and evaluate the practical effectiveness of the resulting recommendation system. The method benchmarks Apriori and FP-Growth across several minimum support values in terms of frequent itemsets count, computation time, and peak memory usage, and then integrates the validated rules into a POS application for real-time inference. The system’s effectiveness is measured using a session-level recommendation acceptance rate, defined as the proportion of transactions that display the recommendation modal and result in at least one recommended item being accepted and paid. The results show that Apriori and FP-Growth consistently produce identical sets of frequent itemsets, but with markedly different computational characteristics: Apriori is significantly faster, while FP-Growth exhibits more stable memory usage. In the deployed setting, the recommendation system achieves a session-level acceptance rate of 15.52% in 3,588 transactions, indicating that roughly one in seven sessions with recommendations leads to an additional item being purchased. Compared to many existing works that focus only on algorithmic performance on benchmark datasets, this research contributes a lifespan-aware, empirically benchmarked hybrid ARM approach that is fully integrated into a working POS system and evaluated using real-world acceptance behavior.

Teknologi: Jurnal Ilmiah Sistem Informasi

Identifikasi Ucapan Disartria Menggunakan Arsitektur Convolutional Neural Network MobileNetV2 dan MFCC

Sistem Rekomendasi Menu Kantin Menggunakan Lifespan-Aware Association Rule Mining Dengan Hybrid Apriori Dan FP-Growth