Global-Local Self-Attention-Based Long Short-Term Memory with Optimization Algorithm for Speaker Identification

Speaker identification (SI) involves recognizing a speaker from a group of unknown speakers, while speaker verification (SV) determines if a given voice sample belongs to a particular person. The main drawbacks of SI are session variability, noise in the background, and insufficient information. To mitigate the limitations mentioned above, this research proposes Global Local Self-Attention (GLSA) based Long Short-Term Memory (LSTM) with Exponential Neighborhood – Grey Wolf Optimization (EN-GWO) method for effective speaker identification using TIMIT and VoxCeleb 1 datasets. The GLSA is incorporated in LSTM, which focuses on the required data, and the hyperparameters are tuned using the EN-GWO, which enhances speaker identification performance. The GLSA-LSTM with EN-GWO method acquires an accuracy of 99.36% on the TIMIT dataset, and an accuracy of 93.45% on the VoxCeleb 1 datasets, while compared to SincNet and Generative Adversarial Network (SincGAN) and Hybrid Neural Network – Support Vector Machine (NN-SVM). ABSTRAK: Pengenalpastian pembicara (Speaker Identification, SI) melibatkan pengenalan pembicara daripada kumpulan pembicara yang tidak dikenali, manakala pengesahan pembicara (Speaker Verification, SV) menentukan sama ada sampel suara tertentu milik seseorang individu. Kekurangan utama dalam SI ialah variasi sesi, bunyi latar belakang, dan maklumat yang tidak mencukupi. Untuk mengatasi kekangan tersebut, kajian ini mencadangkan kaedah Global Local Self-Attention (GLSA) berasaskan Long Short-Term Memory (LSTM) dengan Pengoptimuman Grey Wolf Jiranan Eksponen (EN-GWO) bagi pengenalpastian pembicara yang berkesan menggunakan set data TIMIT dan VoxCeleb 1. GLSA digabungkan dalam LSTM yang memberi tumpuan pada data yang diperlukan, manakala parameter hiper ditala menggunakan EN-GWO untuk meningkatkan prestasi pengenalpastian pembicara. Kaedah GLSA-LSTM dengan EN-GWO mencapai ketepatan 99.36% pada dataset TIMIT dan ketepatan 93.45% pada dataset VoxCeleb 1, berbanding dengan SincNet dan Generative Adversarial Network (SincGAN) serta Hybrid Neural Network – Support Vector Machine (NN-SVM).