http://103.99.128.19:8080/xmlui/handle/123456789/76 Thesis published in Dept. of ETE 2026-10-04T07:31:18Z http://103.99.128.19:8080/xmlui/handle/123456789/596 Improved Speech Enhancement Through Attention-Driven T-F Masking Strategies Akter, Khadija; ID:, 19METE012P Speech enhancement (SE) aims to elevate the perceptual quality and intelligibil ity of speech signals by mitigating ambient noise and distortions. Recent speech enhancement methods struggle with preserving fine temporal details in noisy en vironments, where low-resolution speech increases noise sensitivity. The difficulty of conveying these details, crucial for understanding speech, is compounded by the computational complexity of mapping-based approaches, which require learn ing the full range of clean spectrogram values. However, this work proposes a masking-based strategy to address these challenges in monaural speech.Two dis tinct approaches are explored by employing the Ideal Ratio Mask (IRM) into deep learning frameworks: a U-Net inspired architecture and a Time-Frequency Transformer Network (TF-TransNet). The effectiveness of both approaches is analyzed across different Signal-to-Noise Ratios (SNRs) using several objective metrics: Short-Time Objective Intelligibility (STOI), Perceptual Evaluation of Speech Quality (PESQ), Segmental Signal-to-Noise Ratio (SSNR) and Scale Invariant Signal-to-Distortion Ratio (SI-SDR). In the first approach, the U-Net model comprises an encoder for feature extraction, a decoder for reconstructing the clean speech signal, and skip connections that enable the direct transfer of key information between the encoder and decoder. The U-Net approach demon strates significant improvements in speech intelligibility and quality under mod erate to high SNRs and familiar noise types, although its performance declines in low SNR conditions and with unseen noise types. To address these limita tions, this work introduces the TF-TransNet, aiming to predict the IRM. This network utilizes a Time-Frequency Attention (TFA) encoder to process the noisy magnitude spectrogram, integrating the processed information with the decoder through a Time-Frequency (T-F) Transformer layer. By employing Multi-Head Self-Attention (MHSA) mechanisms and replacing the initial fully connected layer with a Long Short-Term Memory Unit (LSTM) in the T-F transformer layer, the model can dynamically prioritize informative features and their relationships across both time and frequency domains, enhancing its ability to capture long range dependencies in sequential data. The proposed TF-TransNet outperforms existing models, such as the Convolutional Recurrent Network (CRN) and Gated Convolutional Recurrent Network (GCRN), demonstrating enhanced speech in telligibility and quality across both familiar and unfamiliar noise scenarios. A Master of Engineering (M.Engg) Thesis in Electronics and Telecommunication Engineering (ETE) Department at Chittagong University of Engineering and Technology (CUET). 2024-12-12T00:00:00Z http://103.99.128.19:8080/xmlui/handle/123456789/581 Machine Learning Assisted Beam-Steering Microstrip Patch Array Antenna Design Hossain, Md. Farhad; ID:, 19METE025P The rapid evolution of wireless communication technologies, notably the transi tion from 4G to 5G, has driven the need for advanced antenna systems to meet growing demands for seamless connectivity and enhanced performance. This the sis focuses on designing and optimizing beam-steering microstrip patch antennas for 5G communication, with specific emphasis on the sub-6 GHz frequency band and millimeter-wave applications. A comprehensive methodology is presented, in tegrating traditional antenna design principles with machine learning-based pre dictive modeling and optimization techniques. The initial phases of the research involve designing a 2 × 2 microstrip array antenna, simulated for beam-steering applications at 3.5 GHz. A dataset is curated through parametric studies, and eight machine learning models, including Random Forest, Decision Tree Regres sion, XGBoost, and Gaussian Process Regression, have been trained to predict antenna performance metrics such as return loss (S11) and bandwidth. Among these, Decision Tree Regression emerged as the best overall performer, offering the most accurate predictions, while Random Forest also demonstrated close alignment with the actual values. Optimization strategies refined the antenna design by combining machine learning predictions with advanced algorithms like L-BFGS-B, Genetic Algorithms, and Simulated Annealing. The optimized S11 parameter improved significantly, reaching at-51.41 dB compared to the initial-45.21 dB, indicating the superior impedance matching at 3.5 GHz. Similarly, the optimized S22 value reduced to-53.09 dB from-49.57 dB, further enhancing performance. The gain of the antenna increased from 12.82 dBi to 13.58 dBi, and the directivity improved from 12.89 dBi to 13.69 dBi, demonstrating substantial performance enhancements. Port isolation has been maintained with S12 and S21 values below-10 dB. Additionally, the array antenna achieved a beam switching capacity ranging from −17◦ to +17◦, making it highly effective for beam-steering applications. This research contributes to the field by showcasing the potential of machine learning-assisted antenna design and optimization, addressing chal lenges such as computational complexity and design accuracy. The findings hold promise for 5G networks and satellite communications applications, paving the way for more efficient and reliable antenna systems A Master of Science (M.Sc) Thesis in Electronics and Telecommunication Engineering (ETE) Department at Chittagong University of Engineering and Technology (CUET). 2025-01-26T00:00:00Z http://103.99.128.19:8080/xmlui/handle/123456789/574 A HYBRID DEEP LEARNING MODEL FOR HANDWRITTEN BANGLA AND ENGLISH DIGIT RECOGNITION Akbar, Md. Ali; Student ID:, 19METE004F Convolutional Neural Network (CNN) is widely used for handwritten Bangla and English digit recognition. The problem with the conventional CNN is that it has an intricate and time-consuming convolution process to create a feature map and it requires a lot of computation time during the classification phase. This work primarily proposes substituting the convolution kernels of the classic CNN model with dilated convolution kernels to address the limitations in handwritten digit recognition. Although the computation time decreases, the Dilated CNN model’s feature extraction part still performs inefficiently due to information loss, which lowers accuracy. Observing these above-mentioned problems, a Hybrid Dilated CNN (HDC) model is proposed by using dilated convolution kernels with different dilation rates to eliminate the detail loss problem of the Dilated CNN model. The proposed Hybrid Dilated CNN (HDC) model can achieve larger receptive fields and can extract distant features. Finally, the Support Vector Machine (SVM) and K-Nearest Neighbor (KNN) are applied as a classifier to increase the digit classification accuracy. The BanglaLekha-Isolated dataset of Bangla digits is used to verify the proposed methods and demonstrate that under the same conditions, the accuracy of the Dilated CNN, HDC, and the proposed HDC with KNN are 95.85%, 96.72%, and 96.80%, respectively. The proposed HDC model with KNN shows 100% accuracy in the case of the MNIST dataset of English digits. The HDC – KNN model's effectiveness in recognizing digits of both Bangla and English language is perfectly demonstrated by the values of Precision, Recall, and F1 Score parameters. More evidence of the HDC – KNN model's superiority comes from the values of TPR and FPR parameters that are calculated for each of the digit classes in the case of both languages. These experimental findings suggest that integrating the HDC with the KNN algorithm improves the effectiveness of recognizing handwritten Bangla and English digits. A Master of Engineering (M.Engg) Thesis in ETE Department at Chittagong University of Engineering and Technology (CUET). 2024-10-07T00:00:00Z http://103.99.128.19:8080/xmlui/handle/123456789/564 Deep Learning based Bandwidth Extension and Speech Restoration Taher, Taieba; ID:, 19METE011P Speech Bandwidth Extension (SBE) remains a challenging task in speech pro cessing. It involves the intricate process of estimating missing frequency com ponents in low-resolution signals to reconstruct high-resolution transformations. SBE is a technique used to enhance the quality of speech signals by expanding the frequency range of the audio. This endeavor is fundamental for enhancing speech quality, naturalness, and intelligibility, particularly in scenarios marred by background noise and channel distortions. Low-resolution signals often lack high-frequency components due to limited sampling rates and band-limiting, as dictated by the Nyquist-Shannon Sampling Theorem. The inherent loss of high frequency details stems from aliasing, quantization errors, and insufficient granu larity in sampling, which poses significant challenges in restoring full-band speech. Addressing these issues is critical to advancing SBE techniques. Traditional methodologies heavily relied on statistical algorithms and learning techniques such as Multi-Layer Perceptrons (MLP). However, the paradigm shift brought about by Deep Learning has revolutionized this domain, opening avenues for more efficient and effective solutions. This thesis delves into deep learning-based approaches for addressing the complexities of BWE and speech restoration, par ticularly emphasizing noisy environments. This research explores three distinct augmentations of a frequency-domain deep learning network, each tailored to tackle specific challenges encountered in bandwidth extension and speech en hancement. The first approach introduces a joint bandwidth expansion and speech enhancement paradigm utilizing Deep Neural Networks (DNNs). This approach is meticulously designed to simultaneously expand the bandwidth of speech signals, reduce noise interference, and maintain the quality and intelligibil ity of the speech. Leveraging the inherent capabilities of DNNs, this methodology accurately estimates missing speech components. It characterizes noise profiles within degraded signals, synthesizing high-fidelity full-band speech from limited bandwidth inputs. The experimentation demonstrates the superior performance of this DNN-based approach, surpassing conventional methods and presenting promising avenues for real-world applications. Stepping beyond conventional methodologies, the second approach introduces an end-to-end frequency-domain framework, aptly named the Robust extension-plus-enhancement of speech utiliz ing Dual-former Network (RDNet). RDNet represents a paradigm shift in BWE, aiming to recover full-band speech from noisy low-band signals directly. By intri cately integrating speech enhancement and ideal bandwidth extension modules ii 0.0– Abstract within a unified framework, RDNet demonstrates remarkable performance across diverse noisy environments. Leveraging short-time Fourier transform (STFT) fea tures for enhancement, RDNet achieves significant improvements in critical met rics such as Short-Time Objective Intelligibility (STOI), Perceptual Evaluation of Speech Quality (PESQ), and log-spectral distortion (LSD), underscoring its efficacy and potential for practical deployment. Finally, this study’s investigation culminates in introducing the conformer-motivated super-denoised network (CS DNet), a novel approach tailored to mitigate the mismatch problem inherent in domain-specific encoders/decoders. By embedding a conv Unet with T-F trans formation layers, CSDNet reduces dependency on training data and outperforms recent baseline methods across various objective metrics. Moreover, CSDNet ex hibits favorable subjective performance in comparative studies, reaffirming its suitability for real-time applications. Through these meticulously crafted ap proaches, this thesis aims to advance the frontier of deep learning-based speech bandwidth extension and speech restoration. By addressing the core challenges of high-frequency component loss in low-resolution signals and proposing inno vative solutions, this work offers insights and methodologies, proposes practical solutions for tackling real-life problems, and contributes to the wider conversa tion in the fields of speech processing and audio engineering A Master of Science (M.Sc) Thesis in Electronics and Telecommunication Engineering (ETE) Department at Chittagong University of Engineering and Technology (CUET). 2025-03-11T00:00:00Z