Title: Unsupervised VAD method based on short-time energy and spectral centroid in Arabic speech case

Authors: Hind Ait Mait; Noureddine Aboutabit

Addresses: Laboratory LIPIM, National School of Applied Sciences Khouribga, Sultan Moulay Slimane University, Morocco ' Laboratory LIPIM, National School of Applied Sciences Khouribga, Sultan Moulay Slimane University, Morocco

Abstract: Voice Activity Detection (VAD) distinguishes speech segments from noise or silence areas. An efficient and noise-robust VAD system can be widely used for emerging speech technologies such as wireless communication and speech recognition. In this paper, we propose two versions of an unsupervised Arabic VAD method based on the combination of the Short-Time Energy (STE) and the Spectral Centroid (SC) features for formulating a typical threshold to detect speech areas. The first version compares only the STE feature to the threshold (STE-VAD). In contrast, the second compares the SC vector and the threshold (SC-VAD). The two versions of our VAD method were tested on 770 sentences of the Arabphone corpus, which were recorded in clean and noisy environments and evaluated under different values of Signal-to-Noise-Ratio. The experiments demonstrated the robustness of the STE-VAD in terms of accuracy and Mean Square Error.

Keywords: unsupervised VAD; short-time energy; spectral centroid; Arabic speech; computer applications.

DOI: 10.1504/IJCAT.2024.141931

International Journal of Computer Applications in Technology, 2024 Vol.74 No.3, pp.158 - 170

Received: 13 Dec 2022
Accepted: 08 Aug 2023
Published online: 03 Oct 2024 *

Full-text access for editors Full-text access for subscribers Purchase this article Comment on this article