8th International Conference on Spoken Language Processing

Jeju Island, Korea
October 4-8, 2004

Improved Robustness of Time-Frequency Principal Components (TFPC) by Synergy of Methods in Different Domains

Shang-nien Tsai

National Taiwan University, Taiwan

Our previously proposed integration of time-frequency principal components (TFPC) features and histogram equalization (HEQ) has improved the robustness of TFPC features under mismatched conditions. To further enhance the robustness of TFPC features, we herein propose to (a) replace HEQ with another feature normalization technique, progressive histogram equalization (PHEQ), (b) combine a spectral noise reduction method, two-stage Wiener filter, and (c) add a temporal robustness algorithm, SNR-dependent waveform processing, which enhances not only the overall SNR but also the periodicity of noisy speech waveform. Although with the same goal of reducing the performance gap caused by noise, these algorithms mentioned above do operate in different domains, and therefore their synergy significantly improves the robustness of TFPC features. Extensive experiments with respect to AURORA2 database are conducted to verify the effectiveness of each technique, and the overall feature extraction scheme gives a relative error reduction of 25.17 % over our previous proposal.

Full Paper

Bibliographic reference.  Tsai, Shang-nien (2004): "Improved robustness of time-frequency principal components (TFPC) by synergy of methods in different domains", In INTERSPEECH-2004, 977-980.