Analysis of Variational Mode Functions for Robust Detection of Vowels

Surbhi Sakshi, Avinash Kumar, Gayadhar Pradhan


In this work, initially the speech signal is decomposed into variational mode functions (VMFs) with the aid of variational mode decomposition (VMD). Each decomposed VMF represents different frequency band of the input speech signal. An approximate speech signal is then reconstructed by using a set of selected VMFs whose center frequency predominantly corresponds to the frequency range of the vowels. In the reconstructed speech signal, energy due to the high frequency unvoiced sound units and noises is suppressed. Consequently, over an analysis frame, the mean of the square magnitude (MSM) of the sample points is significantly higher for the vowels than other sound units. Further, the MSM at each time instant is non-linearly mapped (NLM) using a negative exponential functions to enhance the transitions at the onset and the offset points of vowels and suppress small fluctuations. The NLM-MSM is used as a front-end feature for discriminating vowels in a given speech signal. The experiments conducted on TIMIT database show that, the proposed approach outperforms the existing methods for the task of detecting vowels in a given speech signal under clean and noisy test scenarios.


 DOI: 10.21437/Interspeech.2018-1947

Cite as: Sakshi, S., Kumar, A., Pradhan, G. (2018) Analysis of Variational Mode Functions for Robust Detection of Vowels. Proc. Interspeech 2018, 756-760, DOI: 10.21437/Interspeech.2018-1947.


@inproceedings{Sakshi2018,
  author={Surbhi Sakshi and Avinash Kumar and Gayadhar Pradhan},
  title={Analysis of Variational Mode Functions for Robust Detection of Vowels},
  year=2018,
  booktitle={Proc. Interspeech 2018},
  pages={756--760},
  doi={10.21437/Interspeech.2018-1947},
  url={http://dx.doi.org/10.21437/Interspeech.2018-1947}
}