This presentation was about exploring self-directed hearing aid fitting with two procedures: (1) preference-based adjustment of hearing aid’s amplification profile, (2) loudness-perception profiling.
Kursun, Bertan & Petersen, Erik & Shen, Yi. (2023). Exploring Self-directed Hearing-aid Fitting with No Booth And No Audiogram. 10.13140/RG.2.2.19575.19360.
Bertan Kursun, Erik Petersen, Yi Shen
Auditory brain stem response (ABR) measurements provide a method to determine the hearing threshold of laboratory animals and humans who are unable to provide behavioral responses to auditory stimuli. Determining hearing threshold using ABR typically requires a large number of pre-selected frequency and stimulus levels, which may not be the most efficient method to sample the stimulus space. The goal of the current study is to develop an adaptive procedure that determines in situ the stimulus that will provide the best estimate of the threshold, based on data collected earlier in the procedure. A Gaussian Process model is iteratively fitted after each test level, and the subsequent test level is chosen at the interim threshold estimate predicted by the current model fit. Simulations of the adaptive procedure were conducted using previously collected toneburst ABR measurements from mice presented at 5 dB increments from 0 to 80 dB at seven frequencies ranging from 4 to 32 kHz. The thresholds determined by this procedure were compared to those produced by human raters who viewed the response at all stimulus levels. Initial results indicate that the threshold may be accurately estimated utilizing fewer stimuli levels than current methods, thereby reducing the duration of the measurements.
10.1121/10.0011243Petersen, Erik & Shen, Yi. (2023). Determining auditory brain stem response threshold using an adaptive stimulus selection procedure. The Journal of the Acoustical Society of America. 151.
Erik Petersen, Yi Shen
The auditory brainstem response (ABR) can be used to estimate the hearing threshold of animals or human subjects who are unable to respond to behavioral measures. However, ABR can be time-intensive and is vulnerable to human subjectivity while interpreting the waveforms. An adaptive procedure has been developed to efficiently and objectively estimate ABR threshold. The procedure iteratively fits a Gaussian process (GP) model to the ABR waveforms collected so far and optimizes the subsequent stimulus. The algorithm was validated through numerical simulation using pre-collected human ABR at 0.5, 1.0, 2.0, and 4.0 kHz with levels from below threshold to 90 dB in 5 dB increments from a cohort of normal-hearing listeners. This led to a full stimulus space of approximately 55 stimuli per test ear. For each test ear, the ABR threshold was estimated by human raters based on the entire stimulus space. The ABR threshold was also estimated using the adaptive procedure, which iteratively sampled a subset (~35%) of the stimulus space. At the end of the adaptive procedure, the fitted GP model was able to capture the individual differences in waveform morphology between listeners. Furthermore, the threshold estimates from repeated runs of the adaptive procedure demonstrated adequate test-retest reliability.
Erik Petersen, Yi Shen
Wide dynamic range compression (WDRC) and noise reduction (NR) algorithms are common hearing aid technologies in digital hearing aids, and since these are nonlinear algorithms, they may cause the output signal-to-noise ratio (SNR) to be different from the input SNR for hearing aids. The current study aims to investigate the effects of temporal features of background noise (i.e., the number of background talkers) and amplification prescriptions, based on audiometric configurations, on SNR at the output of the hearing aid. This is the first study to investigate these effects with a commercial hearing aid with both WDRC and NR algorithms.
Donghyeon Yun, Jennifer Lentz, Yi Shen
Two approaches are examined to quantify the synergetic and redundant interactions across spectral regions for sentence recognition in noise. In one approach, the Speech Intelligibility Index is extended such that speech intelligibility is not only influenced by the speech audibility metrics in individual frequency bands but also their pairwise products to capture the synergetic and redundant interactions between pairs of bands. In a second approach, it is assumed that successful keyword recognition is determined by how many independent channels of information is received by the listener. Each channel of information may be uniquely carried by a frequency band, representing the independent contribution of that band, or shared among multiple bands, representing information redundancy. A previously collected dataset was re-analyzed using the two quantification approaches. For each of the 30 listeners in the dataset, correctness in keyword recognition in the IEEE sentences was available for 600 trials. From trial to trial, the sentence-in-noise stimuli varied in terms of signal-to-noise ratio and which frequency bands were removed from the spectrum via filtering. Both approaches provided consistent depictions of how information is shared across spectral regions for the IEEE sentences. Moreover, the information redundancies across frequency regions did not equally impact all listeners.
Yi Shen
Various methods were developed to allow users to fit their hearing devices on their own without the supervision of a hearing professional. These self-fitting procedures can be categorized into active learning-based [5] and self-adjustment of gain characteristics [2,3]. Our goal is to implement a procedure which combines the benefits of both types [4].
Bertan Kursun, Chemay Rigzin Shola, Lauren Langley, Yi Shen
Two approaches are examined to quantify the synergetic and redundant interactions across spectral regions for sentence recognition in noise. In one approach, the Speech Intelligibility Index is extended such that speech intelligibility is not only influenced by the speech audibility metrics in individual frequency bands but also their pairwise products to capture the synergetic and redundant interactions between pairs of bands. In a second approach, it is assumed that successful keyword recognition is determined by how many independent channels of information is received by the listener. Each channel of information may be uniquely carried by a frequency band, representing the independent contribution of that band, or shared among multiple bands, representing information redundancy. A previously collected dataset was re-analyzed using the two quantification approaches. For each of the 30 listeners in the dataset, correctness in keyword recognition in the IEEE sentences was available for 600 trials. From trial to trial, the sentence-in-noise stimuli varied in terms of signal-to-noise ratio and which frequency bands were removed from the spectrum via filtering. Both approaches provided consistent depictions of how information is shared across spectral regions for the IEEE sentences. Moreover, the information redundancies across frequency regions did not equally impact all listeners.
Yi Shen
To capture a listener’s loudness perception profile, categorical loudness scaling (CLS) is typically repeated at various frequencies. The current study aims to develop psychophysical procedures that enable simultaneous estimation of loudness growth across frequencies. For these procedures, the listener hears a pure-tone stimulus provides a categorical rating (“Soft,” “Loud,” etc.) on each trial. After a response is collected, the procedures update a model of the loudness profile and leverage the model to optimize the stimulus (i.e., level and frequency) for the next trial. The modified slope-adaptive procedure selects the stimulus from a uniform distribution spanning the model-predicted dynamic range, while the modified maximum expected information (MEI) procedure optimizes the stimulus based on an entropy metric. Monte Carlo simulations were conducted to evaluate the two procedures using a database that consists of CLS data collected from 148 listeners at Boys Town National Research Hospital. For each listener, the two procedures were run based on responses simulated using their known loudness profiles (i.e., the ground truth). Both procedures were able to estimate the loudness profile close to the ground truth, with a root-mean-square error (RMSE) of about 6 dB after 100 trials. Below 100 trials, the modified MEI procedure showed a lower RMSE.
Yi Shen, Yihui Zhang, Winnie Shao, Stephen T. Neely
The band importance function that captures how the spectral weight varies across frequencies was estimated for sentence recognition in noises with steady-state or fluctuating temporal envelopes from ten young, normal-hearing adult listeners. The test sentences were from either the IEEE or AzBio corpus. The background noise was a 12-talker babble, either unmodulated or amplitude-modulated using an 8-Hz sinusoidal modulator. In the co-located condition, the noise was presented from the same loudspeaker as the target sentence in front of the listener (0°). In the spatially separated condition, the noise was presented simultaneously from two loudspeakers on either side of the target speaker (±90°). The co-located condition was replicated in a separated test session, at least one week from the first session, which demonstrated the reliability of the estimated band importance functions. With the introduction of amplitude modulation, an increase of spectral weight in the 2-kHz frequency band and a decrease of weight in the 250-Hz band was observed, which was consistent for the two test materials and for the two spatial placements of the noise. Compared to amplitude modulation, spatial separation between the target and noise had a less influence on the band importance function.
Yi Shen, Lauren Langley
In psychoacoustics, data collection from human listeners is commonly done by adaptively varying the stimulus parameters according to the data collected from preceding trials during an auditory task. Compared to the method of constant stimuli, the adaptive procedures allow the adjustment of stimuli for individual listeners to achieve a desired performance range. The adaptive procedures may be designed to estimate performance thresholds, psychometric functions, or the parameters of psychophysical models. The algorithms within these procedures that govern the selection of stimulus parameters may be based on heuristics or Bayesian active learning, and they may differ in terms of computational complexity, accuracy, reliability, stability, and rate of convergence. Choosing an appropriate adaptive procedure for a specific application can be assisted by Monte-Carlo simulations, in which candidate procedures are run on simulated listeners whose responses are governed by ground truth psychometric functions or psychophysical models. The estimated model parameters can be then compared to the ground-truth parameters to evaluate the performance of each adaptive procedure. In this presentation, an introduction of this approach will be provided with examples from psychoacoustics. The analysis techniques that address accuracy, reliability, stability, and rate of convergence will be discussed.
Yi Shen