Detail Cantuman

Noise merupakan masalah signifikan yang menurunkan kualitas dan
kejelasan sinyal audio. Metode konvensional seringkali kesulitan menangani noise yang kompleks, dinamis, dan non-linier. Penelitian ini bertujuan untuk merancang dan mengevaluasi efektivitas model deep learning Convolutional Recurrent Network (CRN) untuk tugas penjernihan ucapan (speech enhancement). Metode penelitian ini menggunakan pendekatan kuantitatif, di mana model CRN yang mengintegrasikan Convolutional Neural Network (CNN) sebagai encoder-decoder dan Gated Recurrent Unit (GRU) untuk pemodelan temporal, dilatih menggunakan pasangan data audio bersih dan bernoise. Sinyal audio diubah ke domain waktufrekuensi menggunakan Short-Time Fourier Transform (STFT) untuk ekstraksi fitur. Hasil eksperimen menunjukkan bahwa model CRN yang diusulkan berhasil secara signifikan, berbeda dengan model baseline yang gagal total dan hanya menghasilkan keheningan. Model yang diusulkan berhasil meningkatkan Signal-to-Noise Ratio (SNR) hingga +2.8 dB dan mencapai skor Perceptual Evaluation of Speech Quality (PESQ) puncak sebesar 1.0385. Hasil ini membuktikan bahwa arsitektur CRN baik untuk mereduksi noise latar belakang, Meningkatkan kualitas dan kejelasan sinyal audio.

Noise is a significant problem that degrades the quality and intelligibility of
audio signals. Conventional methods often struggle to handle noise that is complex, dynamic, and non-linear. This research aims to design and evaluate the effectiveness of a deep learning Convolutional Recurrent Network (CRN) model for the task of speech enhancement. This study employs a quantitative approach, where the CRN model—integrating a Convolutional Neural Network (CNN) as an encoder-decoder and a Gated Recurrent Unit (GRU) for temporal modeling—was trained on pairs of clean and noisy audio data. Audio signals were transformed into the time-frequency domain using the Short-Time Fourier Transform (STFT) for feature extraction. Experimental results demonstrate that the proposed CRN model was highly successful, in stark contrast to a baseline model that failed completely and only produced silence. The proposed model successfully improved the Signalto-Noise Ratio (SNR) by up to +2.8 dB and achieved a peak Perceptual Evaluation of Speech Quality (PESQ) score of 1.0385. These findings prove that the CRN architecture is a robust and effective solution for reducing background noise, significantly improving the quality and clarity of the audio signal.