Citation

Prakash, V. R. and Sasikumar, S. and Bhuvaneswari, Thangavel and G, Krishna Bharath and V S, Srikanth and Ganesan, Balaji (2025) LSTM-Aided Speech Enhancement with Wiener Filter Adaptation & Learned Loss Function. In: 2025 Multimedia University Engineering Conference, MECON 2025, 21 July 2025 - 23 July 2025, Cyberjaya, Malaysia.

Text 77.pdf - Published Version Restricted to Repository staff only Download (1MB)

Abstract

Voice augmentation is the process of strengthening voice signals that have been affected by background noise. This work presents a deep learningbased speech signal augmentation model that is innovative. The proposed model consists of two phases: Training (i) and Testing (ii). In the training phase, the noisy input signal is processed using a Non-negative Matrix Factorization (NMF) to estimate the signal and noise spectrums. The Wiener filter's features are then extracted using the Empirical Mean Decomposition (EMD) method. The Fractional Delta AMS characteristics are retrieved, the denoised signal is obtained from the EMD, and the bark frequency is evaluated. Here, in this paper we’re using “Learning with Learned Loss Function", for improving PESQ score. The key contribution of this study is the precise estimate of the tuning factor η for the Wiener filter using the Long Short-Term Memory (LSTM) model for each input signal. The extracted features (EMD), which were trained on the LSTM using a modified wiener filter to break down the spectral input, provide the denoised improved speech signal. A comparative analysis is carried out between the proposed and existing models with respect to error metrics.

Item Type:	Conference or Workshop Item (Paper)
Uncontrolled Keywords:	Voice augmentation
Subjects:	Q Science > QA Mathematics > QA71-90 Instruments and machines
Divisions:	Faculty of Engineering and Technology (FET)
Depositing User:	Ms Rosnani Abd Wahab
Date Deposited:	19 Mar 2026 00:19
Last Modified:	19 Mar 2026 00:19
URII:	http://shdl.mmu.edu.my/id/eprint/15605

Downloads

Edit (login required)