Pengenalan Perintah Suara Untuk Kendali Robot Berbasis Optimasi MFCC Dan Arsitektur Deep Learning CNN–LSTM

Santoso, Santoso (2026) Pengenalan Perintah Suara Untuk Kendali Robot Berbasis Optimasi MFCC Dan Arsitektur Deep Learning CNN–LSTM. Doctoral thesis, Insitut Teknologi Sepuluh Nopember.

Text 7022202011-Doctoral.pdf - Accepted Version Restricted to Repository staff only Download (1MB) | Request a copy

Abstract

Penelitian ini bertujuan mengembangkan kerangka pengenalan perintah suara robot terintegrasi melalui tiga tahapan: analisis ekstraksi fitur, evaluasi komparatif arsitektur pembelajaran mendalam, dan pengembangan model hibrida untuk klasifikasi perintah suara. Tahap pertama menganalisis Mel-Frequency Cepstral Coefficients (MFCC) untuk memperoleh representasi fitur ucapan yang stabil terhadap variasi tempo, pitch, durasi, resolusi, dan kejelasan pengucapan menggunakan data perekaman mandiri dari 100 penutur untuk 8 kata perintah bahasa Indonesia (3.872 sampel). Tahap kedua membandingkan kinerja arsitektur 1D CNN, 2D CNN, dan 3D CNN menggunakan subset delapan kelas Google Speech Commands v2 yang terdiri atas sekitar 8.000 sampel, pada kondisi eksperimen yang seragam. Tahap ketiga mengusulkan arsitektur parallel CNN–LSTM yang memproses masukan MFCC secara simultan melalui CNN untuk ekstraksi fitur spektral dan LSTM untuk pemodelan dependensi temporal, kemudian menggabungkannya pada tahapklasifikasi. Hasil penelitian menunjukkan MFCC menghasilkan representasi fitur yang konsisten terhadap variasi pengucapan. Pada evaluasi arsitektur CNN, 3D CNN memberikan akurasi tertinggi (89,61%), sedangkan 2D CNN menunjukkan kompromi terbaik antara akurasi dan efisiensi komputasi. Model parallel CNN– LSTM mencapai akurasi 89,17% pada data validasi dan pengujian, serta menunjukkan kinerja lebih baik dibandingkan model serial CNN–LSTM. Penelitian ini menegaskan bahwa integrasi analisis fitur, evaluasi arsitektur, dan model hibrida paralel dapat menghasilkan sistem pengenalan perintah suara robot yang lebih komprehensif.
======================================================================================================================================

This research aims to develop an integrated robot voice command recognition framework through three stages: feature extraction analysis, comparative evaluation of deep learning architectures, and hybrid model development for voice command classification. The first stage analyzes Mel-Frequency Cepstral Coefficients (MFCC) to obtain stable speech feature representations against variations in tempo, pitch,duration, resolution, and pronunciation clarity using a custom recorded dataset from 100 speakers for 8 Indonesian command words (3,872 samples). The second stage compares the performance of 1D CNN, 2D CNN, and 3D CNN architectures using an eight-class subset of Google Speech Commands v2 comprising approximately 8,000 samples under uniform experimental conditions. The third stage proposes a parallel CNN-LSTM architecture that processes MFCC inputs simultaneously through CNN for spectral feature extraction and LSTM for temporal dependency modeling, then fuses them at the classification stage. The results show that MFCC produces consistent feature representations against pronunciation variations. In CNN architecture evaluation, 3D CNN achieves the highest accuracy (89.61%), while 2D CNN offers the best trade-off between accuracy and computational efficiency. The proposed parallel CNNLSTM model achieves 89.17% accuracy on validation and test data, outperforming the serial CNN-LSTM model. This research confirms that the integration of feature analysis, architecture evaluation, and parallel hybrid models can produce a more comprehensive robot voice command recognition system.

Item Type:	Thesis (Doctoral)
Uncontrolled Keywords:	CNN–LSTM, kendali robot, MFCC, pembelajaran mendalam, pengenalan perintah suara, deep learning, robot control, voice command recognition
Subjects:	T Technology > TK Electrical engineering. Electronics Nuclear engineering > TK3070 Automatic control
Divisions:	Faculty of Intelligent Electrical and Informatics Technology (ELECTICS) > Electrical Engineering > 20001-(S3) PhD Thesis
Depositing User:	ansi aflacha
Date Deposited:	13 Aug 2026 03:29
Last Modified:	13 Aug 2026 03:29
URI:	http://repository.its.ac.id/id/eprint/144339

Actions (login required)

View Item