PENGENALAN PERINTAH SUARA BERBASIS HYBRID DEEP LEARNING DENGAN METODE EKSTRAKSI CIRI BERBASIS POWER-LAW

HUZAIFI HAFIZHAHULLAH, . (2023) PENGENALAN PERINTAH SUARA BERBASIS HYBRID DEEP LEARNING DENGAN METODE EKSTRAKSI CIRI BERBASIS POWER-LAW. Sarjana thesis, UNIVERSITAS NEGERI JAKARTA.

	Text COVER.pdf Download (1MB)
	Text BAB 1.pdf Download (143kB)
	Text BAB 2.pdf Restricted to Registered users only Download (794kB) | Request a copy
	Text BAB 3.pdf Restricted to Registered users only Download (549kB) | Request a copy
	Text BAB 4.pdf Restricted to Registered users only Download (1MB) | Request a copy
	Text BAB 5.pdf Restricted to Registered users only Download (123kB) | Request a copy
	Text Daftar Pustaka.pdf Download (230kB)
	Text LAMPIRAN.pdf Restricted to Registered users only Download (3MB) | Request a copy

Abstract

Masalah ketahanan derau masih menjadi hal yang menantang pada sistem pengenalan suara, meskipun kemajuan teknologi deep learning telah digunakan. Kehadiran derau dapat menyebabkan ketidaksesuaian antara pelatihan yang dilakukan dalam kondisi bersih dan kondisi pengujian yang bising. Model deep learning yang banyak digunakan pada pengenalan suara hanya melibatkan model tunggal yang memiliki kemampuan belajar terbatas. Selain itu, fitur berbasis fungsi logaritmik menjadi fitur standar pada banyak sistem pengenalan ucapan. Namun, keandalannya terhadap derau telah menjadi masalah utama. Dalam penelitian ini, penggunaan hybrid deep learning dan ekstraksi ciri berbasis power-law diusulkan. Power-law dapat memberikan kompresi yang lebih baik di daerah berenergi rendah sehingga tidak sensitif ketika sinyal suara terdistorsi oleh derau. Fitur tersebut diimplementasikan pada model dengan menggabungkan dua algoritma deep learning secara paralel, yang selanjutnya disebut dengan hybrid deep learning. Eksperimen ini menggunakan Speech Command Dataset yang disediakan oleh TensorFlow dan dicampur dengan berbagai derau. Hasil eksperimen menunjukkan bahwa penerapan hybrid deep learning dan power-law memperoleh akurasi 84,82% hingga 89,16% dalam hal mengklasifikasi suara berderau. ***** The problem of noise robustness is still a challenge for speech recognition systems, even though advances in deep learning technology have been used. The presence of noise may cause a mismatch between training, which is performed in clean conditions, and noisy testing conditions. The deep learning model that is widely used in voice recognition only involves a single model that has limited learning ability. In addition, features based on logarithmic functions are becoming a standard feature in many speech recognition systems. However, its noise robustness has been a major problem. In this study, the use of hybrid deep learning and power-law based feature extraction is proposed. The power law can provide better compression in low-energy regions so that it is not sensitive when the speech signal is distorted by noise. This feature is implemented in the model by combining two deep learning algorithms in parallel, hereinafter named to as hybrid deep learning. This experiment uses the Speech Command Dataset provided by TensorFlow and mixed with various noises. The experimental results show that the application of hybrid deep learning and power-law obtains an accuracy of 84.82% to 89.16% in terms of classifying noisy sounds.

Item Type:	Thesis (Sarjana)
Additional Information:	1). Dr. rer. nat. Bambang Heru Iswanto, M.Si ; 2). Dr. Eng. Hilman Ferdinandus Pardede, S.T., MEICT
Subjects:	Sains > Matematika > Ilmu Komputer Sains > Fisika
Divisions:	FMIPA > S1 Fisika
Depositing User:	Users 18452 not found.
Date Deposited:	01 Sep 2023 02:26
Last Modified:	01 Sep 2023 02:26
URI:	http://repository.unj.ac.id/id/eprint/39782

Actions (login required)

View Item