Evaluasi Unjuk Kerja Dereverberation Berbasis Machine Learning Terhadap Speech Intelligibility Pengguna Implan Koklea

Asyraf, Muhammad Ammar (2021) Evaluasi Unjuk Kerja Dereverberation Berbasis Machine Learning Terhadap Speech Intelligibility Pengguna Implan Koklea. Undergraduate thesis, Institut Teknologi Sepuluh Nopember.

[img] Text
02311740000018 - Undergraduate_Thesis.pdf - Accepted Version
Restricted to Repository staff only until 1 October 2023.

Download (2MB) | Request a copy

Abstract

Reverberant dan talker masker memberikan pengaruh terhadap performa intelligibility pengguna implan koklea. Evaluasi objektif menunjukkan reverberant dan talker masker secara konsisiten menurunkan intelligibility, kecuali untuk kondisi jumlah talker masker tertentu, Evaluasi subjektif menunjukkan hasil yang berbeda, dimana reverberant pada penggunaan jumlah channel 10 dan 14 memberikan intelligibility yang tinggi pada vocoder CIS dan SPEAK. Fenomena release of masking diperoleh pada banyak kondisi talker masker. Model dereverberation menggunakan metode time-frequency masking digunakan, dengan berbasis recurrent neural network menggunakan fitur audio berupa time-frequency power spectrogram. Model GRU menunjukkan performa optimal dengan waktu komputasi yang lebih cepat dari model LSTM dan SRU. Model dereverberation mampu memberikan peningkatan intelligibility pada tes subjektif sebesar 4% pada vocoder SPEAK dengan 10 channel untuk kondisi reverberant meeting room. Pada analisis objektif, implementasi dereverberation pada kondisi dengan talker masker tidak memberikan indikasi terjadinya fenomena release of masking. ====================================================================== Reverberant and talker masker affect intelligibility of cochleat implan users. Objective evaluastion using STOI shown that these two factors make intelligibility decreased, except for condition dengan certain number of talker masker. Subjective evaluation suggests different result which show reverberant on 10 and 14 channels used in vocoder give high intelligibility near 100% on CIS and SPEAK vocoder. Release of masking phenomenon obtained in many conditions on subjective test. Dereverberation model developed dengan time-frequency masking using recurrent neural network-deep learning model. GRU show optimal performances with fast computation on audio processing, which can improve intelligibility untuk 4% on SPEAK vocoder dengan 10 channels on meeting reverberant condition.

Item Type: Thesis (Undergraduate)
Uncontrolled Keywords: Dereverberation, intelligibility, Time frequency masking, masking release
Subjects: T Technology > TK Electrical engineering. Electronics Nuclear engineering > TK7895.S65 Speech recognition systems
Divisions: Faculty of Industrial Technology and Systems Engineering (INDSYS) > Physics Engineering > 30201-(S1) Undergraduate Thesis
Depositing User: Muhammad Ammar Asyraf
Date Deposited: 06 Sep 2021 07:16
Last Modified: 06 Sep 2021 07:16
URI: https://repository.its.ac.id/id/eprint/91657

Actions (login required)

View Item View Item