Suitability of speech quality evaluation measures in speech enhancement

In this paper, we discuss the suitability of speech quality evaluation measures under various noise environments in the application of spectral subtraction speech enhancement. We take three kinds of typical noise and evaluate comprehensively the speech quality under the standard of global signal-to-noise ratio of noisy speech. We take six kinds of quality measures which include mean opinion score, perceptual evaluation of speech quality, segmental signal-to-noise ratio, weighted spectral slope, log-likelihood ratio and log spectral distance. Then appropriate evaluation algorithms are chosen for speech enhancement based on spectral subtraction. The simulation results show that in the application of speech enhancement, the suitability of speech quality evaluation algorithms is limited to the SNR of noisy speech, recording people, recording content and background noise environment; with regard to three kinds of practical background noise which are taken from TIMIT database, the suitability of the segmental signal-to-noise ratio is the worst and log spectral distance is the best.