The contribution of short-term memory for sound features to speech-in-noise perception and cognition

Speech-in-noise (SIN) perception is a fundamental ability that declines with aging, as does general cognition. We assess whether auditory cognitive ability, in particular short-term memory for sound features, contributes to both. We examined how auditory memory for fundamental sound features, the carrier frequency and amplitude modulation rate of modulated white noise, contributes to SIN perception. We assessed SIN in 153 healthy participants with varying degrees of hearing loss using measures that require single-digit perception (the Digits-in-Noise, DIN) and sentence perception (Speech-in-Babble, SIB). Independent variables were auditory memory and a range of other factors including the Pure Tone Audiogram (PTA), a measure of dichotic pitch-in-noise perception (Huggins pitch), and demographic variables including age and sex. Multiple linear regression models were compared using Bayesian Model Comparison. The best predictor model for DIN included PTA and Huggins pitch (r 2 = 0.32, p < 0.001), whereas the model for SIB included the addition of auditory memory for sound features (r 2 = 0.24, p < 0.001). Further analysis demonstrated that auditory memory also explained a significant portion of the variance (28 %) in scores for a screening cognitive test for dementia. Auditory memory for non-speech sounds may therefore provide an important predictor of both SIN and cognitive ability.


Introduction
As we age, our ability to listen and have conversations in noisy environments, speech-in-noise (SIN) ability, deteriorates.Cognitive abilities also tend to deteriorate.Decline in SIN ability has been linked to age-related cognitive decline and to future risk of dementia (Jiang et al., 2022;Stevenson et al., 2022;McClannahan et al., 2022).Although SIN ability can be accounted for by the degeneration of anatomical structures in the inner ear, the contribution of cognition to successful SIN perception is well recognized (Kujawa and Liberman, 2009;Pronk et al., 2019;Je, 2018).The relationship becomes more important when hearing is impaired and in older populations where phonological working memory is associated with SIN perception ability (Akeroyd, 2008;Füllgrabe and Rosen, 2016).Other work has shown a broad range of cognitive abilities to be associated with SIN ability, including processing speed, inhibitory control, working memory, episodic memory and IQ (Pronk et al., 2019;Dryden et al., 2017).However, these studies use a variety of SIN tests, which makes a direct comparison between studies difficult.
In this study, we assessed whether the contribution of auditory memory for sound features, frequency and amplitude modulation rate, relates to digits-in-noise and sentences-in-noise.Our previous work has shown that auditory memory for frequency correlates with a sentence level measure of SIN ability (Lad et al., 2021;Lad et al., 2020).We hypothesized that auditory memory would be more critical to the sentence-in-noise measure than digits-in-noise measures due to the longer duration of trials in the former which would activate auditory memory.We also assessed the relevance of auditory memory to a screening measure of general cognition, as conventional cognitive testing can be time-consuming and lead to confounding results based on fluency and cultural factors (Nielsen, 2022).This would then support the application of simple automated auditory memory measures to clinical settings as biomarkers for cognitive function and dementia risk.

Participants
153 participants (96 female) were invited to participate in this study from a variety of sources including local participant registries and clinical settings.The age range of participants was 50 -86 years with a mean of 67 years and a standard deviation of 10 years.23 participants were active hearing aid users.Further details are shown in Table 1.Participants using hearing aids did not wear them for audiometric threshold testing but used over-ear headphones for the remainder of auditory testing.These participants did not have any documented evidence of any neurological or psychiatric comorbidity.All participants had the opportunity to go through a Participant Information Leaflet before enrolment in the study.This study was approved by the Oxford C NHS Research Ethics Committee -21/SC/0139.

Auditory and behavioural testing procedure
Each participant had a 1-hour visit to the Auditory Laboratory at the Newcastle University Medical School.Pure tone audiometry (PTA) testing was performed on both ears from 250 Hz to 8 kHz at octave intervals for air conduction using an Interacoustics AS608e screening audiometer in accordance with British Society of Audiology testing guidelines.Pure tones were manually presented as short bursts twice starting at 30 dB HL then increased in 5 dB HL increments until comfortably audible if necessary.Then 5 dB HL reductions were made until the tone was not audible.This process was repeated twice, and the lowest audible volume was chosen as the value for a particular frequency.If maximum amplification at 100 dB HL could not be perceived, then this was used as the ceiling value at a particular frequency.The overall mean of high frequency values between 4 and 8 kHz for the best ear was taken as the threshold value for an individual for further analysis.This value was chosen as high-frequency thresholds are suspected to deteriorate first in age-related hearing loss and previous research from our group has suggested that PTA thresholds in this range correlate with speech-in-noise difficulties (Holmes and Griffiths, 2019;Tl et al., 2008).Hearing status was determined as follows: normal if the mean threshold was below 20 dB HL, mild if between 20 and 40 dB HL, moderate if 40-60 dB HL hearing loss and severe if >80 dB HL.
Computerized auditory testing was performed using Sennheiser HD 201 circumaural headphones.A test of Huggins Pitch perception was performed initially as a test of perception in noise without additional memory requirements (Cramer and Huggins, 1958).The stimulus consists of white noise introduced in one ear whilst the same white noise is phase transformed in a narrow band (6 % around 600 Hz) and played to the other ear.This results in the perception of a faint pure tone, corresponding in pitch to the center frequency of the phase-shifted band, embedded in noise.The 1-sec stimulus was played up to two times for each participant and if they were able to perceive the pitch within the white noise they participated in a short test consisting of six trials where they had to choose the white noise stimulus that had the pitch percept from a set of three sounds played in a sequence with an interstimulus interval of 200 ms.
The digits-in-noise (DIN) task requires participants to listen to three digits on a background of speech-shaped white noise and then select the digits they had heard from a keypad shown on the screen.Participants had two practice trials at the beginning of the task to familiarize themselves with the stimuli at an SNR of 10 dB.An adaptive 1-up, 1down psychophysical paradigm was implemented whereby a correct response resulted in the SNR being reduced and an incorrect one caused the SNR to increase.The starting SNR was 0 dB and the step sizes decreased from 5 to 2 dB after 3 reversals, which then reduced to 0.5 dB after 3 more reversals.The run terminated after 10 reversals and the SNR at the last 5 reversals was averaged to calculate the DIN threshold for each participant.Lower SNR values indicated a better performance.Participants had two practice trials at the beginning of the task to familiarize themselves with the stimuli at an SNR of 10 dB.
The sentence-in-babble (SIB) task required participants to listen to sentences in a background of 16-talker babble as described previously (Lad et al., 2020;Holmes and Griffiths, 2019).Target sentences had the form 〈name〉 〈verb〉 〈number〉 〈adjective〉 <noun> (e.g."Alan gives four pretty flowers") and participants had to click on the correct word from a list of five columns (10 options for each word) shown on the screen with the same structure.The SIB threshold was determined with the same adaptive thresholding procedure followed in the DIN test described above.
Auditory Memory (AuM) was tested using non-speech stimuli as previously described (Fig. 1) (Lad et al., 2021).A one-second tone or AM modulated white noise stimulus was presented to a participant after which they were asked to 'find' the sound on a horizontal scale on a computer screen.Participants had to move a mouse and click on the line to produce a sound at that location.They could make as many clicks as they wanted with no set time limit.After they were satisfied with their choice, they would advance to the next trial by pressing the 'Enter' key on a keyboard.Frequencies that determined the pure-tone sounds were chosen from a uniform distribution between 440 and 880 Hz and AM rates for the white noise stimulus were 5-20 Hz with a sinusoidal function used to apply this modulation.Hanning windows were applied to all synthetic sounds to avoid clicks and the beginning and end of the stimuli.The task consisted of 32 trials with the frequency and AM rate matching trials being interleaved.Participants had a short break after trials.A Gaussian function was used to estimate the standard deviation of the errors in each trial across the whole experiment and the inverse of this value, the precision, was used for further analysis.Thus, one obtains precision scores for frequency AuM (AuM (F)) and AM rate AuM (AuM (A)).Studies in vision have found that this measure better reflects the memory resource a participant can allocate in a given task (Ma et al., 2014).Participants had two practice trials with each stimulus (2 for frequency and 2 for AM rate AuM) at the beginning of the task to familiarize themselves with the stimuli.
Finally, participants completed the short version of the Goldsmiths Musical Sophistication Index (GMSI) questionnaire consisting of questions on paper as a test of musicality (Müllensiefen et al., 2014).Our previous work has found that scores on this test and the AuM measure for frequency precision are associated and the inclusion of GMSI scores allows one to account for variations for AuM (F) performance (Lad et al., 2020).Addenbrooke's Cognitive Examination-Third Edition (ACE-III) was used as the cognitive screening measure (Hsieh et al., 2013).This test is used commonly in memory clinics to screen for cognitive impairment related to neurodegenerative dementias like Alzheimer's disease dementia.

Statistical analysis
Descriptive statistics were calculated using Jupyter Notebooks in Python 3.9.AuM measures were log-transformed to convert them into normally distributed values.All other variables were normally distributed.Bayes Factors were calculated using the Cauchy distribution with a scale parameter of 0.707 as a prior distribution to determine the strength of effect sizes.
First, linear models assessed the effect of age on hearing metrics using the Pingouin module.Multivariate linear regression models to predict DIN and SIB performance were then created using age, PTA thresholds and AuM scores using the SciPy module.Model comparison was performed using Bayesian Information Criterion (BIC) to evaluate the best model.BIC was used instead of the Akaike Information Criterion as the latter tends to overfit the data by using more parameters.To identify the most parsimonious yet informative model, we employed a systematic model refinement process using BIC as a guiding metric.A difference of above 2 between BIC values for models constitutes 'positive' evidence in favor of the model with the lower score.The model with the lowest score was chosen as the 'best' model.P-values of the regression coefficients were used to assess statistical significance at a threshold of p < 0.05 and multicollinearity was assessed using Variance Inflation Factor to remove highly correlated predictor variables which could distort the estimated effects of the model.
Linear models with a combination of Age, PTA thresholds, Sex, Education status, Huggins score, DIN scores, SIB scores, GMSI scores, AuM for frequency performance and AuM for AM rate were assessed for predictability of ACE-3 scores.

Central hearing test performance decline with age
A participant was considered to have passed the Huggins Pitch perception test if they correctly identified the sound that evokes a pitch percept in five or all six trials.103 (67.7 %) participants passed this test.Participants were less likely to pass the test if they were older (χ 2 = 12.4,p = 0.03) or had hearing impairment (χ 2 = 22.1, p < 0.001) (Fig. 3).There were no significant differences between male and female participants.
DIN and SIB thresholds were normally distributed and were transformed into z-scores for further analysis.The mean DIN difference between the target and masker was − 6.1 ± 3.2 dB and 0.2 ± 4.5 dB for the SIB task.There was a weak linear relationship between DIN performance and age after performing a partial correlation with PTA thresholds as a covariate (r = − 0.24 [− 0.4, − 0.08], p < 0.005, BF = 2.39).This was also observed between SIB thresholds and age after using PTA scores as a covariate (r = − 0.19 [− 0.35, − 0.02], p < 0.05, BF = 2.44).
AuM scores were skewed and therefore underwent a logarithmic transformation to normalize the data distribution and a subsequent transformation to z-scores for further analysis.Although both AuM (F) and AuM (A) were both related to age, the latter showed a statistically significant relationship with age (r = − 0.29 [− 0.43, − 0.12], p > 0.05, BF = 2.37) whereas AuM (F) did not show a significant linear relationship (r = − 0.08 [− 0.25, 0.09], p < 0.001) even after adjusting for GMSI scores.

Fig. 1. Auditory Memory
Experiment.An auditory (pure tone or amplitude modulated noise) stimulus is presented for 1 second, then, after a delay of 2 to 4 s, participants can match sounds using a horizontal scale on the screen.The scale is linked to the parameter of interest (frequency for pure tone or AM rate) that can generate the original stimulus after exploring the parameter space to 'find' the stimulus.The figure shows an auditory matching trial where the participant's 'final match' (rightmost dark gray marker on the scale) is shown in comparison to where the original stimulus (orange marker on the scale) is located.In this example, the participant first clicked on the scale to make a 'first match' (which produced a sound linked to the parameter at that location), then a 'second match' and then a 'final match'.The discrepancy between the 'final match' location parameter and that of the original stimulus gives an 'error' for each trial that can be used to calculate the auditory working memory 'precision', the inverse of the standard deviation of errors from a trial target, for all auditory trials.

Fig. 2. A scatterplot showing the relationship between Age and PTA Hearing
Thresholds in Decibels (dB).As age increases, PTA hearing thresholds also increase i.e. the presence of hearing loss increases.Specific gradations of hearing loss severity are indicated by the colors blue (normal), green (mild hearing loss), orange (moderate hearing loss) and red (severe hearing loss).

Best predictors of SIN perception ability
The DIN and SIB tasks differed in the duration of the stimuli and the response time for each trial.SIB matching times (mean = 9.1 s) were significantly longer than DIN matching times by 7.7 (±1.6) seconds (t = 9.5, p < 0.001, BF > 10,000).The longer matching duration is more likely to activate memory resources, therefore, we hypothesized that AuM would better predict SIB performance rather than DIN.
Table 2 shows the statistically significant linear models for predicting DIN scores and their respective BIC scores in ascending order.A linear model with PTA scores and the Huggins pitch score best predicted DIN thresholds with a BIC value of 401.1.The next best statistically significant model had a BIC score of 405.9, more than 4 points larger, and included AuM (F) as an additional predictor variable.The first linear regression model was statistically significant, predicting DIN scores with a r 2 = 0.32, p < 0.001.This relationship is shown in Fig. 4 in a scatterplot.
When SIB thresholds were used as an outcome predicted variable, the best model included PTA thresholds, Huggins pitch scores and AuM (A) scores as predictor variables with a BIC score of 275.6 (Table 3).However, the second-best model including Sex as an additional predictor and the next best model with AuM (F) showed BIC scores of 276.7 and 278.0 respectively.Therefore, predicting SIB, rather than DIN, included additional variables such as AuM(A), Sex and/or AuM(F).The best linear regression model had an r 2 = 0.24, p < 0.001, and is shown in Fig. 4.

Best predictors of cognitive measures
Linear models with a combination of age, PTA thresholds, sex, Education status, Huggins score, DIN scores, SIB scores, GMSI scores, AuM (F) and AuM (A) were assessed for predictability of ACE-3 scores.As

Table 2
The various statistically significant linear models used to predict Digits-in-Noise  above, the best model was determined as having the lowest BIC score by a difference of 2 from the next best model.Table 4 shows the BIC scores of the top 8 performing models that were statistically significant.The best model included the variables AuM (A), SIB and Education status with a BIC score of 745.2.The next best included the addition of DIN scores with a BIC or 750.1.The relationship between AuM (A) scores, adjusted for Education status and SIB, and ACE-3 results showed statistical significance in a linear regression model (r 2 = 0.28, p < 0.001).This is shown in Fig. 5.

Discussion
This study highlights the importance of auditory memory for sound features in SIN perception and general cognition.Auditory memory for temporal fluctuations, as indicated by performance in the AuM (A) task, may be particularly important for sentence-in-noise perception rather than digits-in-noise.Performance on the AuM (A) task was also an important variable in the prediction of cognitive test results in the ACE-3, a commonly used screening neuropsychological test for dementia in memory clinics.This underscores how simple auditory memory measures like auditory memory can potentially be applied to clinical settings as biomarkers for cognitive function and dementia risk in the future.

The relationship between auditory memory and speech-in-noise perception ability
There is a myriad of SIN perception tests used in research and clinical settings with different characteristics.These have shown a differential involvement of auditory memory when predicting SIN ability.
Phonological working memory measures such as the reading span have been shown to predict SIN perception (Akeroyd, 2008).The reading span requires processing and storing several sentences and recalling the last word associated with each sentence; this requires memory over tens of seconds.SIN ability was commonly measured in these tasks using sentence-in-noise perception which may require processing over a length of time where working memory measures and SIN ability would be expected to correlate (Lad et al., 2020).We replicated this finding in the current work with a larger sample size showing that AuM for temporal fluctuations in sound may be better related to SIN perception for sentences than measures related to the carrier frequency.However, models with both variables were statistically significant.
Some studies have noted that auditory memory may only become relevant in older individuals and those with hearing impairment (Füllgrabe and Rosen, 2016).The reasons for this are unclear but may be related to greater resources from working memory being required for understanding sentences in noisy backgrounds when perceptual machinery is partly degenerated.We found that the linear model including PTA thresholds, Huggins pitch scores and AuM (A) was the best predictor of SIB scores showing that all three variables may have independent effects.This may have been due to the nature of the AuM task we used, which was based on resource allocation models rather than phonological measures.The AuM precision measure has been shown to be related to the total memory 'resource' an individual is able to allocate to an item that needs to be remembered and is based on a continuous measure rather than the conventional discrete phonological measures (Ma et al., 2014).This measure may therefore be more sensitive to capture an individual's general memory ability regardless of hearing status or age.

Duration of speech-in-perception tests affects relationship with auditory memory measures
This study suggests that the choice of paradigm used to test SIN ability may determine if auditory memory is needed for the task.Whilst the best model for predicting DIN scores included PTA scores and Huggins scores as predictors, the best model that predicted SIB scores required the addition of AuM precision for AM rate in the linear model.The latter finding is congruent with our previous finding about the relationship between SIB performance and AuM, but the much larger sample size of this study and diverse participant characteristics allows for a nuanced analysis of the data (Lad et al., 2020).That study

Table 3
The statistically significant linear models used to predict Speech-in-Babble (SIB) scores.Bayesian Information Criterion (BIC) scores are shown alongside each model.The lowest score indicates the best suited model for each test (indicated by an asterisk).suggested a better relationship between SIB and AuM precision for the carrier frequency, but the smaller sample size did not allow for the inclusion of covariates to provide an accurate assessment of the data.The current data suggests that during sentence-in-noise perception, memory over a period of tens of seconds for AM stimuli may be a crucial element that allows individuals to communicate successfully.Despite DIN scores being best predicted by Huggins scores and PTA scores only, the second-best model included the addition of AuM for frequency precision.However, the BIC value for this model was more than 4 points greater than the former.This suggests that AuM may still be important for SIN perception at shorter time intervals (the stimulus in the DIN task was presented over three seconds).Further work is necessary to evaluate whether similar results would be found when SIN is tested with word-in-noise tasks where the stimuli only consist of a single word that must be correctly perceived on a noisy background over a shorter duration.At longer time durations (above ten seconds), for the SIB task, other models that were close to the best model consisting of PTA scores, Huggins pitch scores and AuM precision for AM rate included the addition of AuM precision for frequency and sex suggesting that stimulus independent mechanisms of AuM memory may be relevant to SIN perception.This suggests a greater involvement of auditory cognitive factors relating to memory for communicating in noisy environments.There are also linguistic factors, such as syntactic complexity, associated with sentence processing that may be associated with AuM for basic sound features that underlie these relationships (Vogelzang et al., 2021).For example, sentences with a word order that deviate from canonical norms are more effortful to comprehend than those that do not.However, the sentences used in the SIB task had a very simple syntactic structure, which all participants found easy to understand immediately.

How might auditory memory for temporal fluctuations relate to speech-in-noise perception for sentences?
The similarities shared by AM white noise stimuli and speech in natural conversation may explain the significant relationships between AuM and SIN perception ability in our study.Both speech signals and AM noise have temporal fluctuations in their amplitude envelopes (Goswami, 2019).These fluctuations in speech arise due to the natural rhythm, stress patterns, and syllable structure of spoken language.In AM noise, these fluctuations are artificially imposed by modulating the amplitude of the noise over time and this can allow us to study how the auditory system processes and encodes temporal information in speech, even in the absence of background noise.The modulation patterns in AM noise can be designed to resemble the natural modulation patterns found in speech signals by manipulating the modulation rate and depth.In this study, the AM rates used were between five and 20, which is like those present in natural speech.With AM noise, listeners must process and track the signal's temporal fluctuations like tracking the words and gaps between them in sentences.In both cases, cognitive resources such as attention and working memory are engaged, potentially activating similar brain mechanisms.
The AuM precision task for AM rate may capture an individual's general ability to remember speech-like sounds over several seconds, which may be more relevant to SIN perception ability than memory for the carrier frequency.Holding frequency of a sound over time, a property related to the source of a sound, may not confer an advantage for speech intelligibility over memory for AM rate, a property which relates to temporal fluctuations within speech.The latter may be relevant to speech factors that are independent of a speaker such as the envelope of a sound or a sentence.The results for AuM trials for frequency may also be confounded by an individual's ability to rehearse or remember the pitch of the stimulus which may be dependent on factors such as musical sophistication (Kaernbach and Schlemmer, 2008).However, the analysis with the addition of this measure, in the form of GMSI scores, also did not improve the fit of the model with AuM frequency.This may have been due to the shared properties of the AuM (A) and AuM (F) tasks measuring domain-general auditory memory processes rather than a specific advantage for memory for a specific sound feature.There is evidence that AM phase relationships are important in speech rhythm perception, particularly those sounds with strong onsets like 'stress syllables' that produce large amplitude fluctuations in the speech envelope (Leong et al., 2014).Speech also becomes unintelligible if these important fluctuations are removed (Doelling et al., 2014).As mentioned previously, the AM rates used in the experimental task mirrored the rates present in human speech, which may have contributed to its inclusion in the linear models predicting sentence-in-noise perception being better than AuM (F).
There are significant differences between speech stimuli and AM noise that must be noted.Speech stimuli contain linguistic information, phonetic features, and specific spectral characteristics that are not present in AM white noise.One of the differences between the DIN and SIB tasks was the use of numbers in the former and words making up sentences in the latter.The difference in results between the two tests could potentially be attributable to differences in semantic complexity between the stimuli as this was not controlled for (Paivio et al., 1968).However, the fact that the non-speech stimuli used in the AuM tasks were able to capture relationships with verbal SIN tasks shows that there may be general auditory cognitive abilities that are common to both constructs.

Relevance of auditory memory to general cognitive ability measured by screening measures for dementia
Auditory memory is an attractive cognitive mechanism for the link between SIN perception ability and dementia risk due to its underlying neuroanatomical substrates (Griffiths et al., 2020).Auditory memory is increasingly important for SIN perception as one ages and develops hearing loss and is a plausible bridge between SIN perception and dementia (Akeroyd, 2008;Füllgrabe and Rosen, 2016).Previous work from our laboratory has suggested that auditory memory for several seconds is potentially mediated by medial temporal lobe structures.The maintenance of tones, without interference, for more than 3 s activates the hippocampus with fMRI and shows low-frequency oscillatory activity in the region with iEEG (Kumar et al., 2021;Kumar et al., 2016).Interestingly, this study did also show greater activation in the hippocampus for longer maintenance periods potentially reflecting long-term memory resources being utilized.In the task, participants were asked to keep a retro cued pure tone in mind for up to 16 s and choose which tone they had kept in mind at the end of the trial.Evidence from vision suggests that brain structures involved in long-term memory processes can be activated by short-term memory tasks that require high-resolution memory or the binding of multiple visual features or that performance in these tasks can be affected by damage to them (Borders et al., 2022;Yonelinas, 2013;Zokaei et al., 2019).Thus, medial temporal lobe activity may underlie successful auditory memory and SIN perception and manifest as both abilities deteriorating in dementia.
Cross-sectional work has previously demonstrated that auditory cognitive tasks using SIN tasks can be used to identify group level differences between people without dementia, those with mild cognitive impairment and AD dementia (Idrizbegovic et al., 2011).The process of auditory scene analysis, where different elements of auditory input need to be identified, is also impaired in people with AD dementia as compared to those who are cognitively healthy (Goll et al., 2012).This work showed poor group-level performance for people with AD dementia for tasks assessing auditory object segregation and grouping, using synthetic complex sound stimuli below the level of speech, conventionally associated with temporal regions of the brain.The tasks used in these experiments required the tracking of sound features over several seconds, invoking auditory memory, which may be the underlying reasons for these findings in AD patients.In the current study, we did not specifically test associations with the subsections of the ACE-3 due to potential ceiling effects and further work is necessary to test associations after extensive neuropsychological testing using standardized measures for episodic memory, which is one of the key cognitive domains affected in AD dementia.Previous work has suggested that the AuM tasks may also be associated with scores on the forward and backward digit span but whether it also relates to non-verbal measures of attention and working memory has not been explored (Lad et al., 2020).This may also allow us to assess whether there is a stronger relationship with specific domains of cognition, rather than a modest relationship that was observed with the overall score of the ACE-3.
AuM could potentially be used as an important functional measure of cognition in people at risk of dementia.The task has the advantage of being free of linguistic and cultural factors that could potentially influence performance on common cognitive tests, for example.Furthermore, the task can be automated and performed with instructions in the form of a video.This has the added advantage of needing fewer healthcare professionals in the assessment of an individual at risk of dementia.The test could also be used alongside neurodegenerative biomarkers to monitor cognitive impairment due to AD, once confirmed with fluid biomarkers.The combination of neuropsychological testing and such biomarkers has been shown to be predictive of AD dementia with a high degree of accuracy (Palmqvist et al., 2021).

Limitations and future directions
There are multiple factors which can influence hearing beyond those examined in this study.There is inherent variability in genetics and the environment, such as occupational noise exposure, that can affect hearing from the peripheral to the central areas of the auditory system (Gates and Mills, 2005).Humans are also exposed to varying degrees of noise over their lifetime.This study did not include an evaluation of previous noise exposure, like occupational or environmental exposure.Although this study assessed other factors that are evidenced to have a large effect on SIN perception ability, further work needs to consider elements such as genetics, environmental noise exposure and socioeconomic factors as predictor variables (Tsimpida et al., 2019).Another limitation of this study is that the SIN measures included two tests and so we were not able to study the effect of AuM and SIN perception in a parameterized fashion.For example, using measures such as word-in-noise perception ability will allow us to study if linguistic factors require AuM and increasing the digits span to 4 or 5 digits would test whether thresholds with a longer span of digits were predicted by models including AuM (Wilson et al., 2007).Finally, a larger sample size would have allowed methods such as cross-validation and a separate test sample to accurately assess the predictability of the linear models using gold-standard statistical methods.For example, with more than 500 participants, we could have employed 5-fold cross-validation with around 100 participants in every fold.

Conclusion
This work underscores the relevance of auditory processing of basic sound features to a person's SIN perception ability and cognition.This study showed that SIN perception involving sentences and over long durations is best predicted by models involving AuM and the latter was a crucial component of models that best predicted cognition.Further work needs to assess the applicability of such findings to test whether AuM can help predict cognitive decline in dementia.

Fig. 3 .
Fig. 3. Huggins Pitch test performance.Bar chart showing the percentage of people passing the test per decade (x-axis) (top image) and the bottom image shows this percentage by hearing loss severity.

Fig. 4 .
Fig. 4. The scatterplot in the top panel shows a statistically significant linear relationship (r 2 = 0.32, p < 0.001) between normalized PTA threshold scores, adjusted for Huggins pitch scores, and normalized DIN thresholds.The scatterplot in the bottom panel shows a significant linear relationship (r 2 = 0.24, p < 0.001) between normalized AuM (A) scores, adjusted for PTA scores and Education, with standardized SIB scores.

Fig. 5 .
Fig. 5.The scatterplot shows a statistically significant linear relationship (r 2 = 0.28, p < 0.001) between normalized AuM scores for amplitude modulation rate precision, adjusted for Education and SIB scores, and cognitive scores in the form of ACE-3 results.

Table 1 Demographic, hearing and cognitive testing scores are shown for all partici- pants. PTA scores are split by hearing status and sex. Standard deviations are presented in parentheses for all metrics apart from the count making up par- ticipants with different levels of hearing loss where the percentage proportion is indicated.
PTA -Pure Tone Audiogram scores, AuM (F) -Auditory Memory precision for Frequency, AuM (A) -Auditory Memory precision for AM Rate, GMSI -Goldsmiths Musical Sophistication Index score.Education Status: 1-left school at 15 years of age, 2-completed schooling/1 professional qualification, 3-multiple professional degrees.

Table 4
The various linear models used to predict ACE-3 scores.Bayesian Information Criterion (BIC) scores are shown alongside each model.The models are ranked from best to worst performing.The lowest score indicates the best suited model for each test (indicated by an asterisk).