ISCA Archive Interspeech 2020
ISCA Archive Interspeech 2020

A Noise Robust Technique for Detecting Vowels in Speech Signals

Avinash Kumar, S. Shahnawazuddin, Waquar Ahmad

In this work, we propose a novel and noise robust method for the detection of vowels in speech signals. The proposed approach combines variational mode decomposition (VMD) and non-local means (NLM) estimation for the detection of vowels in a speech sequence. The VMD algorithm is used to determine a number of variational mode functions (VMFs). The lower-order VMFs represent the frequency contents corresponding to vowel regions. Thus by combining the lower-order VMFs and reconstructing the speech signal back, the energy corresponding to the vowel regions is enhanced while the non-vowel regions are suppressed. At the same time, the ill-effect of noise is also reduced. Finally, as reported in an earlier work, application of NLM followed by convolution with first-order difference of Gaussian window is performed on the reconstructed signal to determine the vowel region. The performance of proposed approach for the task of detecting vowels in speech is compared with three existing techniques and observed to be superior under clean as well as noisy test conditions.