Extension and evaluation of a near-end listening enhancement algorithm for listeners with normal and impaired hearing

Abstract

In many applications in which speech is played back via a sound reinforcement system such as public address systems and mobile phones, speech intelligibility is degraded by additive environmental noise. A possible solution to maintain high intelligibility in noise is to pre-process the speech signal based on the estimated noise power at the position of the listener. The previously proposed AdaptDRC algorithm [Schepker, Rennies, and Doclo (2015). J. Acoust. Soc. Am. 138, 2692-2706] applies both frequency shaping and dynamic range compression under an equal-power constraint, where the processing is adaptively controlled by short-term estimates of the speech intelligibility index. Previous evaluations of the algorithm have focused on normal-hearing listeners. In this study, the algorithm was extended with an adaptive gain stage under an equal-peak-power constraint, and evaluated with eleven normal-hearing and ten mildly to moderately hearing-impaired listeners. For normal-hearing listeners, average improvements in speech reception thresholds of about 4 and 8 dB compared to the unprocessed reference condition were measured for the original algorithm and its extension, respectively. For hearing-impaired listeners, the average improvements were about 2 and 6 dB, indicating that the relative improvement due to the proposed adaptive gain stage was larger for these listeners than the benefit of the original processing stages.

Speech Modifications to Increase the Intelligibility of Vocal Messages Broadcast by Driving Assistance Systems Intended For Hearing-Impaired Drivers

“…For high SII values, the speech signal is only slightly modified because the intelligibility in the time window is already high whereas for low SII values the signal undergoes a greater modification. The evaluation of these modifications at a constant SPL level compared with the original speech signals on hearing-impaired subjects in the case of a masking noise recorded in a cafeteria show a reduction in SRT of 1.5 dB on average [19]. Nathwani et al [20] used a speech modifications algorithm based on the reproduction of three Lombard effect components to increase the intelligibility of a speech signal in an incar noise.…”