Assessing speech intelligibility in hearing impaired listeners

Quantitatively assessing the speech intelligibility deficits observed in hearing impaired (HI) listeners is a basic component for a better understanding of these deficits and a crucial component for the development of successful compensation strategies. This dissertation describes two main streams of work aiming at a better quantitative understanding: (i) Chapter 2 focused on describing a new analysis framework based on a confusion entropy and a distance metric to analyze consonant-vowel (CV) perception in HI listeners across different listening conditions; (ii) Chapters 3, 4, and 5 focused on developing a speech intelligibility (SI) model to account for observed deficits in HI listeners.

In Chapter 2, HI listeners were provided with two different amplification schemes to help them recognize CVs. In the first experiment, a frequency- independent amplification (flat-gain) was provided. In the second experiment, a frequency-dependent prescriptive gain was provided. An entropy measure and an angular distance measure were proposed to assess the highly individual effects of the frequency-dependent gain on the consonant confusions in the HI listeners.

These measures along with a standard analysis of the recognition scores suggested that, while the average recognition error score obtained with the frequency-dependent amplification was lower than that obtained with the flat-gain, the main confusions made by the listeners on a token basis remained the same in a majority of the cases.

Chapter 3 describes the introduction of the HI deficits of reduced audibility and decreased frequency selectivity into a speech-intelligibility model for normal-hearing (NH) listeners. The NH model is based on a signal-to-noise ratio measure in the envelope domain (SNRe n v ), as presented in the framework of the speech-based envelope power spectrum model (sEPSM, Jørgensen and Dau, 2011; Jørgensen et al., 2013).

The predictions of the model were compared to data in three different noise maskers. While the model was able to account for the relative difference of the HI listeners performance in these different noise interferers; it faild to account for the absolute performance in the noise interferers. Chapter 4 replaced the linear peripheral model, i.e. the gammatone filterbank, by a nonlinear auditory nerve model.

The SI predictions showed good agreement with human data when the model operated at an overall presentation level (OAL) of 50 dB sound pressure level (SPL) and with only medium-spontaneous-rate fibers. However, when all fiber types and a realistic OAL of 65 dB SPL were considered, the model overestimated SI in conditions with modulated noise interferers.

In Chapter 5, the front-end processing of an auditory-nerve (AN) model was combined with a correlation-based back end inspired by the vowel-coding hypothesis of stable rate patterns in the inferior colliculus. The proposed model assesses the correlation between the noisy speech and the noise alone, as represented by the AN model’s bandpass-filtered instantaneous firing rates, assuming an inverse relationship with SI.

The NH listeners’ SI data were accurately predicted for all noise types, additionally demonstrating reasonable changes across presentation levels. Furthermore, the SI for 13 HI subjects was predicted by adjusting the front end parameters specifying the inner and outer hair-cell loss based on the audiogram of the listeners.

The predictions showed good agreement with the measured data for four out of the thirteen subjects and reasonable agreement for a total of eight subjects. The work provides a foundation for quantitatively modeling individual effects of inner and outer hair-cell loss on speech intelligibility.