A Filter Constructed from Estimation of Clean Speech and Noise for Speech Enhancement in Speech Recognition Systems

Author

Meng Sha ; Qin Shenghao ; Liu Jia

Author_Institution

Dept. of Electron. Eng., Tsinghua Univ., Beijing

fYear

2006

fDate

4-6 Oct. 2006

Firstpage

1620

Lastpage

1623

Abstract

In this paper, a filter used for speech enhancement in speech recognition systems is presented. The filter is based on the Mel-warped Wiener filter and model-based compensation for additive noise. The filter coefficients are designed in Mel-spectral energy domain and then transformed to impulse response in time-domain using Mel-warped inverse discrete cosine transform (Mel-IDCT). The enhanced speech signal is finally obtained by applying the impulse response in time-domain. The performance of the proposed filter has been tested using 863 Database under different noise conditions. The result shows that the speech enhanced by the proposed filter has much higher word accuracy in speech recognition systems

Keywords

Wiener filters; discrete cosine transforms; speech enhancement; speech recognition; time-domain analysis; transient response; Mel-spectral energy domain; Mel-warped Wiener filter; Mel-warped inverse discrete cosine transform; additive noise; impulse response; robust ASR; speech enhancement; speech recognition systems; speech signal; Additive noise; Automatic speech recognition; Degradation; Noise robustness; Speech enhancement; Speech recognition; Systems engineering and theory; Time domain analysis; Wiener filter; Working environment noise; Mel-warped Wiener filter; Model-based compensation; Robust ASR; Speech enhancement;

fLanguage

English

Publisher

ieee

Conference_Titel

Computational Engineering in Systems Applications, IMACS Multiconference on

Conference_Location

Beijing

Print_ISBN

7-302-13922-9

Electronic_ISBN

7-900718-14-1

Type

conf

DOI

10.1109/CESA.2006.313573

Filename

4105639