Informatics Publishing Limited

Mourad Talbi ¹, Anis Ben Aicha ², Lotfi Salhi ¹, Adnane Cherif ³

Affiliations
1 Unit Research of Signal Processing, Science of Tunis, TN
2 Unit Research Techtra, Higher School of Communication of Tunis, University of Carthage, El Ghazala City, TN
3 Unit Research of Signal Processing, Science of Tunis, IN

Abstract

In this paper, we have proposed a new speech enhancement technique based on the time adaptation of stationary bionic wavelet thresholds. The time dependence is introduced by approximating the Teager Energy (TE) of the stationary bionic wavelet coefficients. The Stationary Bionic Wavelet Transform (SBWT) was used in this work, in order to have a perfect reconstruction of the speech signal. Experimental results based on Mean Square Error  (MSE) between the original speech signal and the reconstructed speech signal, show that the SBWT is better than the Bionic Wavelet Transform (BWT). This fact can improve the speech enhancement system based on thresholding in the bionic wavelet domain. We have also used the pause detection to introduce more improvement of the noise level estimation in the stationary bionic wavelet subbands. We have also used the time-space adapted thresholds in order to improve the procedure of stationary bionic wavelet thresholding and avoiding severs degradations of the enhanced speech signal. The proposed technique was evaluated by comparing it to others techniques such as the technique based on Wiener filtering and the techniques based on thresholding in wavelet and bionic wavelet domains and the technique based on undecimated wavelet packet-perceptual filterbanks and MMSE-STSA estimation. The two thresholding techniques are based on time adaptation of wavelet thresholds too. The obtained results from the SNR, SSNR, ISd and PESQ computation show that our proposed technique outperforms the others techniques.

Keywords

Speech Enhancement, Stationary Bionic Wavelet Transform, Teager Energy Operator, Thresholding.

Full Text

     

Aymen Mnassri ¹, Mohammed Bennasr ¹, Adnane Cherif ¹

Affiliations
1 Laboratory of Analysis and Processing of Signals and Electric and Energy System, FST Tunis – 2092, TN

Abstract

Objectives: This paper proposes a novel recognition technique (ASR) based on GA optimized SVM multi-class algorithm. Methods/Statistical Analysis: The Kernel parameters of support vector machine are very important problems that have a great influence on the performance of recognition rate. Thus, GA is adapted to optimize the penalty parameter C and the kernel parameter λ for SVM multi-class, which leads to improve classification performance. Finally, the proposed model is tested experimentally using eleven Arabic words mono-locator. Each word of them is extracted by Mel Frequency Cepstral Coefficients (MFCCs) and used as an input to the SVM multi-class classifier. Findings: The proposed method enhances the recognition rate which is performed to 100% within short duration training time. Application/Improvements: The obtained results shows that the GA-SVM technique achieved the better performance

Keywords

Automatic Speech Recognition, Genetic Algorithm, Mel Frequency Cepstrum Coefficients, Supports Vector Machines

Full Text