The graphs in Figure 1 show the average hearing threshold at 500, 1000, and 2000 Hz (top) and the average hearing threshold at 2000-8000 Hz (bottom) versus scores on a clinical speech-in-noise test know as the Quick SIN. The Quick SIN measures the speech-to-noise (SNR) ratio required to receive 50% words correct on a sentence test compared to data obtained from normal hearing persons. The resulting score is a SNR difference between the hearing-impaired score and the normal hearing score, or SNR loss. As can be readily seen, there is little relation between the average hearing loss and a listeners ability to recognize speech in noise.

In order to better understand the variability across hearing-impaired listeners to understand speech in noise we have begun to develop models or simulations, of auditory signals processed through hearing-impaired ears. The models use mathematical formulations of auditory processing to transform sound into dynamic time-frequency fluctuations thought to be important to the brain when decoding and classifying speech sounds into meaningful language units. Model parameters are adjusted using data obtained in separate experiments on individual patients to reflect elevated auditory thresholds, as well as suprathreshold distortions related to loudness recruitment and reduced spectral resolution.

Loudness recruitment refers to an abnormally rapid growth of perceived loudness as the input signal is increased in level. Hearing-impaired individuals are thought to have a more rapid growth of loudness than normal hearing individuals. We assess this aspect of auditory distortion by a measure of cochlear compression, or the amount of output gain at different frequencies as a function of input level. This measure tells us how much gain the cochlear provides to weak, medium, and high-level input signals.

Spectral resolution refers to how well listeners can process different sounds presented simultaneously to different frequency regions. For example, if you are played a low tone near middle C (approximately 256 Hz) at a loud but not uncomfortable level, are you able to detect a very weak higher tone about a half-octave up around 362 Hz? If you have normal hearing, the level of the 362 Hz tone required to just be detectable would be about the same regardless of whether the tone at middle C was there or not. However, for many hearing-impaired listeners, the presence of the tone at middle C will have an impact on the ability to hear the 362 Hz tone, and the higher tone will have to be made louder than if it were played alone in order for it to be detected. In auditory modeling, the early stages of processing are often thought of as a number of separate channels or filters much like the keys along a piano. The shape and sharpness of the filters determines the number of distinct keys from low to high frequency. We can estimate the shape of these channels by asking normal hearing and hearing-impaired patients to detect tones of different frequencies in the presence of narrow bands of noise positioned both lower and higher in frequency than the target tone. By measuring the interference of these noise bands on tones at different frequencies and at different levels, its possible to estimate the shape of the filter at the tone frequency and how this shape changes with level.

For each listener tested, we constructed a separate auditory model incorporating threshold, filter shape, and amplitude compression. We also constructed a model representing normal hearing with data obtained from a group of normal hearing individuals. By playing speech signals and noise to the different hearing-impaired models and comparing the model outputs to that of the normal auditory model, we are able to quantify differences between what speech might look like to the normal ear/brain versus the impaired ear/brain. These differences are then converted into a score that can be used to predict speech recognition, and the predicted scores are compared to actual scores.

 

 

Figure 2 shows the modeling results for five patients listening to speech in noise at a SNR of -3 dB. The top panel shows results as a function of a simple model of audibility alone (Speech Transmission Index). The bottom panel shows the results of including suprathreshold distortion factors as well as audibility. Each model score is a comparison between speech and noise passed through a simulation of the individuals distorted auditory system and speech alone passed through a simulation of a normal ear. The figure shows that we can account for the differences across hearing-impaired listeners to a much greater extent than when only threshold data are considered. This work highlights the need to incorporate factors related to suprathreshold distortion in order to predict more accurately speech understanding in noise. From a practical point of view, these models, if proven to be robust across a larger number of hearing impaired, can serve as a tool for predicting performance and for developing new signal processing strategies for advanced hearing aids that attempt to compensate for suprathreshold distortions. Conceptually, this is similar to what hearing aids do now by amplifying different frequencies depending on the degree of hearing loss and by applying amplitude compression to compensate for reduced dynamic range. Because the types of suprathreshold distortion are very complicated and can change dynamically with input level, the kinds of signal processing anticipated are likely to be quite complex. Our efforts to characterize the suprathreshold distortion in individual listeners and to embed them into a model of auditory processing that includes peripheral (cochlear) and central (auditory cortex) representations have demonstrated feasibility of this concept. Future work to include additional measures of hearing capacity, such as pitch perception, will likely improve our efforts to model the suprathreshold distortions introduced by impaired hearing.