Predicting Speech Intelligibility Using a Nonlinear and Level-Dependent Auditory Processing Front End

Relaño-Iborra et al. [2016, J. Acoust. Soc. Am., 140(4), 2670-2679] proposed a model, termed sEPSMcorr, which showed that the correlation between the envelope representations of clean and degraded speech is a powerful predictor of speech intelligibility in a wide range of listening conditions. However, due to its simplistic linear preprocessing, sEPSMcorr cannot account for the level-dependent effects and nonlinear properties of the sound transduction in the auditory periphery, which is a prerequisite for accounting for the consequences of sensorineural hearing loss.

Thus, in the present study, a more realistic, nonlinear preprocessing was combined with the correlation-based back end. Specifically, the front end of the computational auditory signal processing and perception model [CASP; Jepsen et al. (2008), J. Acoust. Soc. Am. 124(1), 422-438] was employed, which has been shown to successfully account for psychoacoustic data in conditions of, e.g., spectral masking, amplitude-modulation detection as well as forward masking, for both normal-hearing (NH) and hearing impaired listeners.

The proposed speech-based CASP model, denoted sCASP, receives the clean and degraded speech signals as input. The signals are processed through outer- and middle-ear filtering, a nonlinear auditory filterbank including inner- and outer hair-cell processing, adaptation, as well as a modulation filterbank.

The internal representations at the output of these stages are analyzed using a correlation-based back end. Speech intelligibility predictions obtained with the speech-based CASP implementation are presented and compared to NH listener data obtained in conditions of additive noise, phase jitter, ideal binary mask processing and reverberation.

The results demonstrate a large predictive power of the model. As the front end of sCASP can - unlike the front end of its predecessor sEPSMcorr- be parametrized to account for sensorineural hearing loss, the proposed framework may provide a valuable basis for evaluating the consequences of different aspects of hearing loss on speech intelligibility in the various experimental conditions