Voice Activity Detection Algorithm Using Spectral-Correlation and Wavelet-Packet Transformation
- Авторы: Korniienko O.1, Machusky E.1
-
Учреждения:
- National Technical University of Ukraine “Igor Sikorsky Kyiv Polytechnic Institute”
- Выпуск: Том 61, № 5 (2018)
- Страницы: 185-193
- Раздел: Article
- URL: https://journals.rcsi.science/0735-2727/article/view/177205
- DOI: https://doi.org/10.3103/S0735272718050011
- ID: 177205
Цитировать
Аннотация
It is developed the voice activity detection algorithm using noise classification technique. It is proposed the spectral-correlation and wavelet-packet (WP) features of frames for voice activity estimation. There are tested three WP trees for effective representing of audio segments: mel-scaled wavelet packet tree, bark-scaled wavelet packet tree and ERB-scaled (equivalent rectangular bandwidth) wavelet packet tree. Application only two principal components of WP features allows to classify accurately the environment noise. The using wavelet-packet tree design which follows the concept of equivalent rectangular bandwidth for acoustic feature extraction allows to increase the voice/silence segments classification accuracy by at least 4% in compare to other classification based voice activity detection algorithms for different noise.
Об авторах
O. Korniienko
National Technical University of Ukraine “Igor Sikorsky Kyiv Polytechnic Institute”
Автор, ответственный за переписку.
Email: olexandr.korniienko@gmail.com
Украина, Kyiv
E. Machusky
National Technical University of Ukraine “Igor Sikorsky Kyiv Polytechnic Institute”
Email: olexandr.korniienko@gmail.com
Украина, Kyiv