Adaptive Syllable Training Improves Phoneme Identification in Older Listeners with and without Hearing Loss

Abstract

Acoustic-phonetic speech training mitigates confusion between consonants and improves phoneme identification in noise. A novel training paradigm addressed two principles of perceptual learning. First, training benefits are often specific to the trained material; therefore, stimulus variability was reduced by training small sets of phonetically similar consonant-vowel-consonant syllables. Second, the training is most efficient at an optimal difficulty level; accordingly, the noise level was adapted to the participant's competency. Fifty-two adults aged between sixty and ninety years with normal hearing or moderate hearing loss participated in five training sessions within two weeks. Training sets of phonetically similar syllables contained voiced and voiceless stop and fricative consonants, as well as voiced nasals and liquids. Listeners identified consonants at the onset or the coda syllable position by matching the syllables with their orthographic equivalent within a closed set of three alternative symbols. The noise level was adjusted in a staircase procedure. Pre-post-training benefits were quantified as increased accuracy and a decrease in the required signal-to-noise ratio (SNR) and analyzed with regard to the stimulus sets and the participant's hearing abilities. The adaptive training was feasible for older adults with various degrees of hearing loss. Normal-hearing listeners performed with high accuracy at lower SNR after the training. Participants with hearing loss improved consonant accuracy but still required a high SNR. Phoneme identification improved for all stimulus sets. However, syllables within a set required noticeably different SNRs. Most significant gains occurred for voiced and voiceless stop and (af)fricative consonants. The training was beneficial for difficult consonants, but the easiest to identify consonants improved most prominently. The training enabled older listeners with different capabilities to train and improve at an individual 'edge of competence'.

Keywords: adaptive training; aging; auditory rehabilitation; hearing loss; phoneme identification; speech-in-noise perception.

Figure 1

Stimuli and training paradigm. (a): Time series of the syllable ‘Maat’ and the concurrent noise show the temporal relationships between stimuli and noise. The spectrogram of the syllable sound illustrates the distinct spectral patterns during the syllable onset, the vowel, and the coda. (b): Time series and spectrogram for the syllable ‘Maash’ show the different spectral pattern of the coda compared to ‘Maat’ (c): Power spectrum of the speech-like filtered Gaussian noise. (d): Flow chart of one trial of the training procedure. The example shown is from a coda syllable set. Participants had the options to request a stimulus repetition or selected the written syllable best matching with the sound stimulus. Visual feedback was provided and the SNR for the next trial was adjusted according to the participant’s response.

Figure 2

Hearing thresholds and speech-in-noise (SIN) loss. (a): Mean audiograms for participants between their 60s and 90s, averaged across right and left ears. The shaded areas indicate the standard error of the mean. (b): Distribution of participants across the ages. (c): Individual pure-tone-average (PTA) hearing loss versus age, modelled by linear regression. The grey shaded area indicates the confidence interval for the linear model. (d): Individual SIN loss obtained with QuickSIN versus age. (e): SIN loss versus PTA. The dashed lines mark the ranges for three PTA subgroups.

Figure 3

Outcome measures of the syllable training. (a): Example of the course of SNR level during test blocks with the consonant sets C1 and C2. The horizontal lines indicate the mean SNR across all trials. (b): An accuracy measure was obtained from confusion matrices. The example shows the empirical response probabilities of an individual participant. Blue numbers in the bottom row are the sum within the column (stimulus probability = 1.0). Blue numbers in the right column are the sum of response probabilities across the column. (c): Color map of the confusion matrix. (d): Visualization of SNR and d-prime before and after the training of a single point in the d-prime versus SNR data plane. Participants with low d-prime pre-training increased their accuracy after the training while the SNR remained high. Participants with high pre-training d-prime maintained the high accuracy level while decreasing the SNR. The box plots on the right indicate d-prime changes and the upper box plots SNR changes after the training. (e): A gain measure was obtained as the first principal component of the normalized d-prime and SNR differences between the pre- and post-training tests.

Figure 4

Grand mean d-prime and SNR changes across the five training sessions and between pre- and post-training tests. F-statistic and p-values are the main effects of the session determined by ANOVA with repeated measures. Error bars indicate the 95% confidence intervals. Note the different y axis ranges for the training and tests.

Figure 5

Grand mean training benefits for the three syllable sets and phoneme identification at the onset and coda. (a): pre–post-training changes in d-prime; (b): SNR, (c): normalized gain obtained from combining the training-related changes in d-prime and SNR into a single principal component. Error bars indicate the 95% confidence intervals.

Figure 6

Group-mean d-prime for the individual consonants of the C1, C2, and C3 sets for the syllable test before and after training. The asterisks denote significant pre–post-training differences *: p < 0.05, **: p < 0.01, ***: p < 0.001 (paired t-test, FDR corrected); error bars indicate the 95% confidence intervals.

Figure 7

Relationship between syllable recognition and hearing loss. (a): Individual accuracy measured with d-prime pre- and post-training phonemes at the coda position of syllables from the C1 set. Blue and red lines indicate linear models for the relationship between d-prime and PTA. The shaded areas are the 95% confidence regions for the linear model. (b): Individual SNR values for pre–post-training tests and linear modelling with PTA. (c): Group mean d-prime measures for the six stimuli in three subgroups low-PTA (n = 13 lower quartile of PTA), mid-PTA (n = 26 centre quartiles of PTA), and high-PTA (n = 13, upper quartile of PTA). The shaded areas indicate the training-related increase in d-prime. Error bars indicate the 95% confidence intervals. (d): Group mean SNR measures for the six stimuli.