e-ISSN:0976-5166
p-ISSN:2231-3850


INDIAN JOURNAL OF COMPUTER SCIENCE AND ENGINEERING

Call for Papers 2020

Feb 2021 - Volume 12, Issue 1
Deadline: 15 Jan 2021
Publication: 20 Feb 2021

Apr 2021 - Volume 12, Issue 2
Deadline: 15 Mar 2021
Publication: 20 Apr 2021

More

Indexed in

IJCSE Indexed in Scopus

ABSTRACT

Title : FRAMEWORK BASED SUPERVISED VOICE ACTIVITY DETECTION USING LINEAR AND NON-LINEAR FEATURES
Authors : G.B.Gour, Dr. V.Udayashankara, Dr. Dinesh K Badakh, Dr. Yogesh A. Kulkarni
Keywords : Voice activity detection; Pathological voice analysis; Vocal system; Non-linear dynamic methods; SVM.
Issue Date : Nov-Dec 2020
Abstract :
Voice activity detection (VAD) methods based on linear features are limited by linearity assumption and correct estimation of the pitch. The nonlinear dynamic methods can analyse irregular vocal cord behaviours and are found to be useful in the areas of voice study, clinical treatment evaluation, voice classification as per the Titze. The development of VAD using nonlinear features has required more attention. However, speech recordings in most of the practical scenarios like, in hospitals, research centers, video conferencing, and forensics are affected by babble noise. Moreover, correct estimation of signal to noise ratio (SNR) also depends on reliable voice activity detection. By looking at such challenges, the paper presents a framework based VAD using the combination of linear and nonlinear features with the two-step noise reduction (TSNR) for the possible speech enhancement. Speech segments are classified by using a supervised support vector machine (SVM). The framework is evaluated and compared with different time and frequency domain based VADs on the speech containing continuous and sustained vowels with babble noise. For this purpose, the experimental study is carried on continuous NOIZEUS corpus with babble noise at varying SNR levels. As the vocal disorder is more prominent, laryngeal pathologies based database from Saarbruecken Voice Database (SVD) and Laryngeal cancer data are used for sustained vowels. The study revealed the importance of bio-inspired linear and nonlinear features in the VAD. Finally, the proposed VAD is found to be more robust as far as vowels and continued speech are concerned.
Page(s) : 935-942
ISSN : 0976-5166
Source : Vol. 11, No.6
PDF : Download
DOI : 10.21817/indjcse/2020/v11i6/201106181