Supplementary schemes to spectral subtraction for speech enhancement

  • Authors:

  • Hwai-Tsu Hu;Fang-Jang Kuo;Hsin-Jen Wang

  • Affiliations:

  • Department of Electronic Engineering, National I-Lan Institute of Technology, No.1, Shern Nong Road, Taiwan I-Lan 260, Taiwan, ROC;Department of Electronic Engineering, National I-Lan Institute of Technology, No.1, Shern Nong Road, Taiwan I-Lan 260, Taiwan, ROC;Department of Electronic Engineering, National I-Lan Institute of Technology, No.1, Shern Nong Road, Taiwan I-Lan 260, Taiwan, ROC

  • Venue:
  • Speech Communication
  • Year:
  • 2002

Quantified Score

Hi-index 0.00

Visualization

Abstract

Three supplementary schemes, namely spectral smoothing, formant intensification and comb filtering, are proposed to enhance noisy speech that was dealt with a classic spectral subtraction method. Performance improvements are mainly achieved by exploring cepstral properties and comb filtering, and they are evaluated using both objective and subjective measures based on 12 noise-corrupted speech files. Noise types considered in this study consist of white Gaussian and car noise with signal-to-noise ratio (SNR) set to 5 and 10 dB, respectively. Although experimental results indicate certain inconsistency between objective and subjective measures, it can still be concluded that spectral smoothing and comb filtering contribute to the melioration of speech quality for speech corrupted by additive white Gaussian noise. On the other hand, advantages due to formant intensification can only be recognized whenever music noise is suppressed and the original speech formant characteristics are preserved. The employment of all proposed schemes together generally leads to a notable improvement in perceived quality.