Reasons why Current Speech-Enhancement Algorithms do not Improve Speech Intelligibility and Suggested Solutions

Abstract

Existing speech enhancement algorithms can improve speech quality but not speech intelligibility, and the reasons for that are unclear. In the present paper, we present a theoretical framework that can be used to analyze potential factors that can influence the intelligibility of processed speech. More specifically, this framework focuses on the fine-grain analysis of the distortions introduced by speech enhancement algorithms. It is hypothesized that if these distortions are properly controlled, then large gains in intelligibility can be achieved. To test this hypothesis, intelligibility tests are conducted with human listeners in which we present processed speech with controlled speech distortions. The aim of these tests is to assess the perceptual effect of the various distortions that can be introduced by speech enhancement algorithms on speech intelligibility. Results with three different enhancement algorithms indicated that certain distortions are more detrimental to speech intelligibility degradation than others. When these distortions were properly controlled, however, large gains in intelligibility were obtained by human listeners, even by spectral-subtractive algorithms which are known to degrade speech quality and intelligibility.

Percepção de fala no ruído em idosos usuários de próteses auditivas com diferentes microfones e algoritmo de redução de ruído

“…Although studies (12,13,14) have reported that the noise reduction algorithm does not provide improved speech intelligibility and that it may bring, as a result, distortion in the speech signals (3) , these are findings that differ from the ones mentioned in this research. Even if, by reducing interference from dominant noise to the total signal, noise reduction decreases the gain in frequencies on which is not dominant, occurring good chance of losing information that contributes to the intelligibility (3) , it was observed that, in this research, it has positively contributed to the increase of high scores, providing better performance, especially when the noise was facing the individual.…”

Extending the articulation index to account for non-linear distortions introduced by noise-suppression algorithms

“…(4) to be always smaller than S, and thus consider in the computation of the proposed measure only bands in whichŜ < S. The implicit hypothesis is that those bands will contribute more to intelligibility and should thus be included. This was confirmed in listening studies (Loizou and Kim, 2011) in which normal-hearing listeners were presented with speech synthesized to contain either target attenuation distortions alone (i.e., bands withŜ < S) or target amplification distortions alone (i.e., bands withŜ > S). Speech synthesized to contain only target attenuation was always more intelligible, and in fact, it was found to be more intelligible than either the un-processed (noise corrupted) or processed (via the noise-reduction algorithm) speech.…”

“…(3)] are different. One strategy for improving the overall SNR (defined as the weighted sum of SNRs across all bands) is to discard bands with unfavorable (extremely low) SNRs while retaining bands with favorable SNR (see proof in Loizou and Kim, 2011). Such an approach was taken in our prior study and has been shown to improve speech intelligibility by normal-hearing listeners (Kim et al, 2009) as well as by cochlear implant listeners (Hu and Loizou, 2008).…”

Speech distortion weighted multi-channel Wiener filter and its application to speech recognition

“…Specifically, ¼ 2 6 and ¼ 2 3 yield the best WERs at 0, 5, 10 dBs and at 15 dB, respectively, for the proposed filter and rank-1 GEVD-SDW-MWF. The aggressive noise reduction strategy has been also reported to be preferred in terms of improving speech intelligibility in noise [8,9].…”