The role of working memory in speech recognition by hearing-impaired older listeners: Does the task matter?

Abstract

Objective:

Working memory refers to a cognitive system that holds a limited amount of information in a temporarily heightened state of availability, for use in ongoing cognitive tasks. Research suggests a link between working memory and speech recognition. In this study, we investigated this relationship using two working memory tests that differed in regard to the operationalization of the link between working memory and attention: the auditory visual divided attention test (AVDAT) and the widely used reading span test.

Design:

The relationship between speech-in-noise recognition and working memory was examined for two different working memory tests that varied in methodological and theoretical aspects, using a within-subject design.

Study sample:

Nineteen hearing-impaired older listeners participated.

Results:

We found a strong link between the reading span test and speech-in-noise recognition and a less robust link between the AVDAT and speech-in-noise recognition. There was evidence for the role of selective attention in speech-in-noise recognition, shown via the new AVDAT measure.

Conclusion:

Our findings suggest that the strength of the relationship between speech-in-noise recognition and working memory may be influenced by the match between the demands and the stimuli of the speech-in-noise task and those of the working memory test.

Introduction

Working memory and speech recognition

Working memory (WM), one of the most studied cognitive constructs across various disciplines, has been defined and conceptualized in several ways (see Baddeley, 2012 and Cowan, 2017 for reviews). A general definition of WM that has been deemed applicable across different theories of WM and a wide range of implementations of the concept refers to WM as a system of components that holds a limited amount of information in a temporarily heightened state of availability, for use in ongoing information processing tasks (Cowan, 2017).

WM capacity has been linked to speech recognition, particularly in adverse listening conditions, such as in the presence of noise and/or hearing loss (e.g., Souza & Arehart 2015; Strori, Bradlow, & Souza, 2020; Zekveld, Rudner, Johnsrude, & Rönnberg (2013); for reviews, see Akeroyd, 2008; Besser, Koelewijn, Zekveld, Kramer, & Festen, 2013; Souza, Arehart, & Neher, 2015). The Ease of Language Understanding (ELU) model developed by Rönnberg and colleagues (2013) offers a comprehensive description of the relationship between WM capacity and speech recognition. To provide a brief overview, in the ELU model lexical retrieval is facilitated by an unambiguous match between language input and the respective phonological representation stored in long-term memory, with retrieval in an automatic and relatively effortless manner. When the incoming language input is degraded (e.g., by background noise and/or hearing loss), lexical retrieval is impaired by the difficulty in matching the new information to the corresponding phonological representation(s). Consequently, WM is explicitly engaged to facilitate the match. This view has been supported by several studies that found a more robust relationship between WM capacity and speech in noise than speech in quiet and for older adults with hearing impairment (see Akeroyd, 2008; Besser et al., 2013, for reviews) compared to young adults without hearing impairment (Füllgrabe & Rosen, 2016).

Working memory and attention

Recognizing speech in realistic situations, such as in the presence of noise, requires the listener to process a rapidly incoming auditory stream, attend to the relevant part of this stream (speech) while ignoring the irrelevant background noise, concurrently extract information and store that information for integration with subsequent input and later retrieval. It is thus reasonable to expect that speech recognition draws upon both WM and attention resources. More specifically, selective attention (the ability to direct attention to the relevant information and ignore co-occurring irrelevant information in the background), divided attention (the ability to attend to two or more streams of information), and the ability to temporarily manipulate and store task-relevant information in WM will all impact how speech is processed and recognized. WM and attention are considered to be closely linked by a broad consensus in the literature surrounding these multi-faceted constructs (e.g.; Cowan, 1998; Kane, Bleckley, Conway, & Engle, 2001). One predominant view of attention is that of a limited resource for information processing (Wickens, 1980). According to theories that link WM to attention, the limited capacity of WM reflects a limited cognitive resource, which also serves functions typically attributed to attention. The link between WM and attention can be conceptualized in several ways that differ in terms of the functions that draw on the limited attentional resource (see Oberauer, 2019 for a detailed treatment of this topic). Here we focus on two conceptualizations that are relevant for the purposes of the WM measures used in the present study: (1) attention as a limited resource for storage and processing of information (e.g., Daneman & Carpenter, 1980) and (2) attention as a limited resource for controlling purposes (Schneider & Shiffrin, 1977).

In the first conceptualization, attention is shared between “storage” and “processing” task demands. That is, the same attentional resource is required to keep representations available in WM and to carry out other cognitive processes, such as judging the plausibility of a sentence or selecting a response to a stimulus. A central assumption of this view is that attention-demanding cognitive processes/tasks compete with concurrent storage demands. The second conceptualization is referred to as controlled attention, where a central assumption is that the process of controlling the allocation of attention consumes the limited resource, rather than the process of attending to an object/task per se (Shiffrin & Schneider, 1977). Contrary to the first view, controlled attention assumes that the limited attentional resource is needed for the control of what we attend to, not for keeping representations of objects and events in WM.

These conceptualizations have different implications in regard to WM. Specifically, in a situation in which WM receives both relevant and irrelevant information, according to the storage and processing view, attention limits the amount of information that can be retained in WM, not the extent to which the irrelevant information is kept out of WM (i.e., the filtering efficiency or the ratio of relevant-to-irrelevant stimuli in WM). Consequently, individuals with lower WM capacity retain a smaller amount of both relevant and irrelevant information, but the filtering efficiency is independent of WM capacity. In contrast, the controlled-attention view assumes that the limited attentional resource determines the filtering efficiency. Therefore, individuals with lower WM capacity retain the same amount of information as those with higher capacity, but different WM capacities reflect differences in the filtering efficiency.

Measures of working memory and attention in speech recognition

In speech recognition and hearing science research, WM has been predominantly measured by complex span tasks that require the participant to simultaneously process/manipulate and store information for later recall. One of the most widely used complex span tests is the reading span test (RST) developed by Rönnberg and colleagues (adapted from Daneman and Carpenter, 1980). The participant reads sentences on a computer monitor, presented in lists of varying size, one at a time. After reading each sentence, the participant makes a semantic judgment on the sentence while concurrently trying to retain its first and/or last words. At the end of a list of sentences, the participant is prompted to recall the test items (the first and/or last words in the sentences). The load of the task is controlled by gradually increasing the number of sentences in a recall list. The score of the test (an estimation of WM capacity) is the percentage of correctly recalled target words. In terms of the theoretical framework concerning the WM and attention link, the RST (and its variants) incorporate the storage and processing view of this link.

The popularity of the RST in the speech and hearing literature may be attributed to the number of supporting studies that found a relationship between individual scores on this test and speech recognition performance in older listeners with hearing loss. However, there is also evidence indicating the absence of such a link (e.g., Desjardins & Doherty, 2014; see Souza et al., 2015 for a comprehensive review). In addition to the mixed results in the literature, broadly speaking, the dominant use of a single test of WM, such as the RST, may be constraining. From a methodological perspective, no single task can be deemed a perfect or pure measure of a cognitive construct (Conway et al., 2005). From a theoretical perspective, it means the examination of only one of the conceptualizations of the cognitive construct in question. Only a limited number of studies of older hearing-impaired listeners have used multiple tests to either derive a composite/weighted score of capacity from several similar complex span tests (Ng & Rönnberg, 2019; Nagaraj, 2017), or to compare the efficacy of individual complex span tests that are largely similar (e.g., different versions of the RST in Souza & Arehart, 2015).

Additionally, unlike the storage and processing view represented by the widely used complex span tests, the controlled attention view of this link has received limited attention in the speech recognition and hearing literature. A study by Meister et al (2013) found reduced performance in older compared to young adults for speech recognition tasks in a multi-talker setting that required divided attention, and a strong relationship between performance in speech tasks requiring selective attention and working memory capacity. More recently, Gallun and colleagues developed a new measure of WM, the Auditory Visual Divided Attention Test (AVDAT) (Gallun & Jakien, 2019). This new test was adapted from measures originally developed by Cowan and colleagues (Cowan, Fristoe, Elliott, Brunner, & Saults, 2006), where WM is operationalized as depending on the selective attention system (controlled attention), as proposed by Cowan (1998)¹. That is, the test combines the storage aspect of WM with the control of attention. The AVDAT involves several separate components that are categorized as single or dual modality. The two single modality components involve either auditory (lists of digits) or visual stimuli (list of letters), and can be categorized as simple span tasks of WM. The four dual modality components involve the concurrent presentation of both auditory and visual stimuli (lists of digits and letters). In all the components, the task is to store and recall a list of stimuli (auditory or visual) and the task load is controlled by gradually increasing the size of the recall list. The two types of the dual modality components differ in terms of whether the response list is cued (the participant knows in advance which modality list will be reported and can selectively attend to it while ignoring the other), or uncued (the participant does not know in advance which modality will be reported and has to divide attention between the two modalities). Gallun and Jakien (2019) examined the relationship between performance on the AVDAT and speech-on-speech recognition in a complex auditory environment that involved competing talkers in either the same or different locations as the target speech. They found that performance on the AVDAT was correlated with speech performance, albeit, different components of the test predicted performance in different speech task conditions. Specifically, the dual modality component with the cued visual response (a selective attention component) was a significant predictor of speech performance (represented as target-to-masker ratio) in the separated speech condition and of spatial release (the difference between performances in the separated and co-located conditions). The dual modality component with the uncued visual response (a divided attention component) was a significant predictor of speech performance in the co-located speech condition.

Given the importance of WM and attention in speech recognition, a WM test that incorporates separate measures of selective and divided attention can be a useful tool for tackling the role of individual abilities related to these constructs in speech recognition in adverse listening situations. AVDAT is a promising tool in this respect. However, further study is needed to better understand and consolidate it. In addition, the considerations above display the need for studies that implement in tandem different tests of WM that incorporate different theoretical conceptualizations of the link between WM and attention.

The current study

The present study examined the relationship between speech-in-noise recognition and WM capacity in older adults with hearing loss using a new WM test (AVDAT) and the widely-used reading span test (RST). The most crucial methodological contrast regards the secondary “processing” task that requires processing of the incoming stimuli beyond merely attending to them. Namely, the RST, a complex WM span test, involves such a secondary task (semantic judgement of sentences), whereas the AVDAT does not. In the dual modality components of the AVDAT, the only task is to solely attend to either one (cued tasks) or both (uncued tasks) of the two concurrently presented stimuli lists, for recall at the end of the presentation. Table 1 displays the characteristics of the two WM tests.

Table 1.

Overview of the characteristics of the two cognitive tests used in the present study.

Test	Modality	Stimuli	Recall Item	Processing/Secondary Task	Attention-WM link
RST	Visual	Short, grammatical sentences varying in plausibility and presented in blocks of increasing size	First or last words (in each sentence block)	Semantic judgement on the sentence	Storage and processing
AVDAT	Auditory and visual	Digits presented aurally and letters presented visually as lists of increasing length	Sequences of digits (audio) or letters (visual) in the correct order	No processing task	Storage and control of attention

Our goal was to examine the relationship between speech recognition and different WM tests that include methodological and theoretical contrasts concerning the link between WM and attention. We anticipated better speech-in-noise recognition performance to be related to higher scores on the RST, in line with existing literature (e.g., Souza & Arehart, 2015). Regarding the novel AVDAT measure of working memory, given that understanding speech in noise can be assumed to engage both selective and divided attention, we hypothesized that there would be a link between the cued (selective attention) and uncued (divided attention) dual-modality components of the AVDAT and speech recognition, regardless of the test modality. More specifically, we reasoned that the ability to selectively attend to and process target speech while ignoring concurrent background noise would likely be related to the ability to selectively direct attention to the relevant information (the cued recall list) and store it for later retrieval, while ignoring the irrelevant, competing information (performance on the cued dual modality AVDAT tasks). Similarly, the ability to recognize/process and retain each incoming word in a sentence in order to retrieve it at the end of the sentence (speech recognition performance) may be related to the ability to divide attention between two concurrent sources of information and store this information for later retrieval (performance in the uncued dual modality AVDAT tasks). In regard to the relationship between the two WM measures, we may anticipate a correlation between the RST and the dual modality components of the AVDAT if both tests tap into WM to a similar extent. However, given the methodological and theoretical differences that the RST and the dual modality components of the AVDAT incorporate, they may tap into different WM and related cognitive mechanisms (including attention), in which case a weaker correlation between them could be expected. To assess other factors which might affect the relationship between WM and speech-in-noise recognition in our participant sample, we also included the measure of peripheral hearing loss.

Materials and methods

Participants

Participants included 19 adults (11 female) aged 63–89 years (mean age = 73.4 years) with symmetrical sensorineural loss (Figure 1). Nine participants wore hearing aids bilaterally. The mean pure tone average measured at 0.5, 1 and 2 kHz was 35.53 dB (range: 11.67–66.67 dB) in the right ear and 34.12 dB (range: 15–65 dB ) in the left ear. The mean word recognition in quiet scores were 96% (range 66–100%) in the right ear and 95% (range 70–100%) in the left ear. All listeners passed the MoCA cognitive screening test (Nasreddine, Philips et al., 2005), scoring at least 23 out of 30 points with a group mean score of 26.6 (range 23–30). The inclusion of participants scoring 23 or higher is below the originally proposed passing score of 26, but aligns with work demonstrating good sensitivity and specificity for patients with broad sociodemographic backgrounds and/or having a hearing loss (e.g., Luis, Keegan, & Mullen, 2009; Shen, Sherman & Souza, 2020). All listeners were native speakers of American English and reported normal or corrected-to-normal vision. All listeners were compensated at an hourly rate for their time.

Figure 1.

Audiograms of both ears for the participants (N = 19). Individual thresholds: thin lines, and group average: thick bold lines. Right ear is indicated by the ‘o’ marker and left ear by the ‘x’ marker.

Tests and procedure

The present data consisted of two different WM tests along with audiometric results and one speech-in-noise test. Each measure is described in detail below.

Working memory

Reading Span Test (RST).

The abbreviated English-language version of the Reading Span Test, developed by Rönnberg and colleagues (Ng, Rudner, Lunner, Pedersen, & Rönnberg, 2013) was delivered to the participants. This test involves information processing (semantic judgment) and information storage (recall). The stimuli of the test consist of short sentences that are all grammatically correct, but can be semantically plausible or implausible (e.g., “The captain saw his boat” [plausible], “The train sang a song” [implausible]). Participants were asked to read sentences on a computer screen, which appeared one word or word pair at a time. Words or word pairs were presented at a rate of 0.8 seconds per word or word pair, with an interstimulus interval of 75 ms. At the conclusion of each sentence, participants judged whether the sentence made semantic sense by replying “yes” for plausible and “no” for implausible sentences. Two lists each of 2, 3, 4 and 5 sentences were presented in ascending order of length. At the end of each list, participants were randomly queried to recall either the first or the last words from the list of sentences, in any order. The assignment of “first versus last” word recall was randomized across participants. Participants completed a practice list of 2 sentences before moving to the experimental lists. The experimenter recorded the correctly repeated words on a printed form of the visually presented test. The percentage of correctly recalled words (out of the total number of target words in the 28 total sentences) was taken as the measure of WM capacity.

Auditory Visual Divided Attention Test (AVDAT).

The materials for this test consisted of auditory (digits) and visual stimuli (letters). Participants completed single-and dual-modality span tasks where they were asked to report in the order presented a list of digits presented aurally via insert earphones and/or letters presented visually on a computer located in front of them, as described in detail below. The tasks were completed in the following order: visual letter span, followed by the auditory digit span task, followed by the dual-modality task. All tasks were implemented on a Windows computer using MATLAB R2018b software.

Visual letter span task

The stimuli consisted of the letters A, C, E, F, H, I, L, O, and R that were presented in 90-point font at a distance of approximately 48 cm from the participant and at a rate of one letter per second. The visual stimuli appeared in white in the center of a black background, following an orienting stimulus (fixation cross) that appeared for 3 seconds before the trial began. Every participant received the same sequence of letter lists, beginning with three three-item practice lists and proceeding to the test that included two lists per length, starting with three-item lists and increasing by one item at a time to a maximum of nine items. The test ended if both lists of a particular length were recalled incorrectly. For each list, the stimuli that made up the standard sequence had been drawn randomly, without replacement. After the presentation of each list, the participant was asked to recall the list of letters in the order presented and enter their responses via a graphical user interface (GUI) that displayed keypads showing letters and numbers (all presented in 90 point font). Each item selected as a response appeared at the top of the screen. Participants used a mouse to select their responses. The score represented the average sequence length correctly recalled and was calculated by averaging the total number of correctly recalled items across the total number of lists presented.

Auditory digit span task

The auditory stimuli involved digitized recordings of the digits 1–9 spoken by a male talker, which were time-compressed or expanded in order to have a duration of exactly 500 ms and presented at a rate of 1 sec per item. All auditory measures were delivered without hearing aids, via ER-2 insert earphones at a root-mean-square (RMS) level of 40 dB above the Speech Reception Threshold (SRT) measured for spondaic words presented in quiet. This presentation level never exceeded 80 dB SPL. The auditory digit span task was administered and scored in the same fashion as the visual digit span task.

Dual-modality memory task

This task involved the synchronous presentation of auditory and visual lists. There were three different types of trials that were distinguished by the stimulus arrangement and task instructions. In the first type of trial (attend auditory), a cue appeared (a picture of an ear) which indicated to participants that they had to listen carefully to a digit list and ignore a synchronous list of letters that were visually presented. In the second type of trial (attend visual), a cue appeared (a picture of an eye) which indicated to participants that they had to attend to a list of letters that were visually presented and ignore a list of spoken digits presented synchronously. In the third type of trial (attend unknown) a cue appeared (a question mark) which indicated to participants that they were to attend to both the visual and auditory lists. After presentation of the lists, participants were prompted to recall one of the lists, in the order presented. Dual-modality conditions were thus divided into those in which participants knew in advance the modality that they would report and as such, could selectively attend that modality and those in which participants were not informed in advance which modality they would be asked to report and as such, had to attend to divide their attention to both modalities concurrently. The four conditions were presented in a randomly interleaved fashion, the trial types were randomly determined and an equal number of each cue type was presented for each task at each list length. Each list length was presented twice, starting with the 3-item lists and increasing by one item at a time to a maximum of 7-item lists. Similar to the single-modality tasks, the score was calculated by averaging the total number of correctly recalled items across the total number of lists presented.

Speech-in-noise recognition.

Speech-in-noise recognition was measured using the QuickSIN (Killion, Niquette, Gudmundsen, Revit, & Banerjee, 2004) administered binaurally via insert (ER-3A) earphones. The test required the participant to repeat back sentences spoken by a female talker and played in four-talker babble (three males, one female) background noise. The sentences are low-context and each one includes five key words. The sentences were presented in lists of six, each one at a signal-to-noise ratios ranging from +25 dB (first sentence) to 0 dB (last sentence) in 5 dB-steps. Three lists were administered to each participant, one practice and two test lists. The first list presented served as a practice list to familiarize the participant with the task and allow for speech level adjustment. The test was recorded on a compact disc and routed through an Interacoustics AC 40 audiometer (all the stimuli were preloaded on to the audiometer). Speech presentation levels were fixed and were adjusted based on the participant’s hearing loss. In line with the protocol for the QuickSIN, speech levels were set to 70 dB HL for most listeners (with a pure-tone average of 45 dB or lower²), and to a “ loud but ok “ level for listeners with a pure-tone average greater than 45 dB in either ear. The level was decreased to 65 dB HL for five participants and to 60 dB HL for one participant. The test score represents the signal-to-noise ratio required for the listener to repeat 50% of the words correctly. The final score consisted of the average of the two test list scores.

Results

The data were analyzed in the R environment (R Core Team, 2021, version 4.0.4). Participants’ mean speech in noise score measured via the QuickSIN was 3.87 dB SNR loss (Standard Deviation (SD) = 1.80; range 1.5–8 dB). The mean amount of hearing loss across both ears measured by the pure tone average at 0.5, 1, and 2 kHz was 34.74 dB (SD: 12.34 dB; range: 11.67 – 65 dB). Table 2 displays the mean scores and corresponding standard deviations for each WM test/test component.

Table 2.

Mean, standard deviation (SD) and the range of scores across participants (N = 19) in each WM test/test component.

		Single Modality			Dual Modality
	RST (% of correctly recalled words)	Auditory (SM-A) (average digit sequence length)	Visual (SM-V) (average letter sequence length)	Auditory Cued (DM-AC) (average digit sequence length)	Visual Cued (DM-VC) (average letter sequence length)	Auditory Uncued (DM-AU) (average digit sequence length)	Visual Uncued (DM-VU) (average letter sequence length)
Mean	41.67	4.10	3.66	3.69	2.94	2.85	1.57
SD	12.04	0.43	0.43	0.68	0.94	0.65	0.70
Range	21.43 – 64.29	3.38 – 4.75	2.88 – 4.5	2.5 – 4.8	0.1 – 4.3	1.6 – 3.9	0.4 – 2.6

The mean, standard deviation and range of WM scores measured via the reading span test (RST) were consistent with previous studies of similar-aged groups (e.., Souza & Arehart, 2015). For the AVDAT scores, the mean values, standard deviation and ranges for the single- and dual-modality components were relatively similar to the corresponding values in Gallun and Jakien (2019), with some slight differences that may be attributed to the differences in the age range and population samples in the two studies (only older listeners in our study compared to both younger and older adults in the reference study).

Correlations

A correlation analysis was performed to assess the relationship between WM capacity measured by the reading span test (RST) and each of the components of the audio-visual divided attention test (AVDAT), speech-in-noise scores, hearing and age (results are displayed in Table 3). Normality checks, conducted on each variable via Shapiro-Wilks tests, revealed that all but one of the variables (namely, DM-VC) were normally distributed. Pearson correlation was implemented for the normally distributed variables and Spearman correlation for tests involving the non-normally distributed variable. As observed in Table 3, participants’ QuickSIN scores and their performance on the RST were significantly correlated (r =- .54, p = .02), whilst no significant correlations were found between QuickSIN and performance on any of the task components of the AVDAT. Further, no significant correlations were found between the RST score and any of the scores of the selective, or divided attention components of the AVDAT, revealing a weak relationship between performances in these two WM tests.

Table 3.

Correlations between the variables of interest: speech-in-noise recognition (QuickSIN scores), the two different WM measures (RST and the individual components of the AVDAT), hearing (PTA), and age.

	RST	DM-AU	DM-VU	DM-AC	DM-VC	SM-A	SM-V	PTA	Age
QuickSIN	−.54*	−.23	−.27	−.40	−.24	−.13	−.42	.59**	.31
RST		.19	.13	.17	−.02	.33	−.13	−.50*	−.55*
DM-AU			.06	.33	.07	−.09	.31	−.13	−.23
DM-VU				.47*	.46*	.32	.66**	−.37	−.26
DM-AC					.47*	.52*	.53*	−.03	−.14
DM-VC						.20	.53*	−.32	.12
SM-A							.32	−.31	−.08
SM-V								−.31	−.02
PTA									.27

RST: Reading span test; DM-AU: Dual modality task with the auditory stimuli as recall target, not cued; DM-VU: Dual modality task with the visual stimuli as recall target, not cued; DM-AC: Dual modality task with the auditory stimuli as recall target, cued; DM-VC: Dual modality task with the visual stimuli as recall target, cued; SM-A: Single modality task, auditory digit stimuli (average digit span); SM-V: Single modality task, visual letter stimuli (average letter span); PTA: Across-ears pure tone average measured at 0.5, 1 and 2 kHz. Spearman correlation used for tests including the non-normally distributed variable DM-VC.

^*p < .05

^**p < .01, p-values not corrected for multiple comparisons.

Linear regression analysis

Speech-in-noise scores were analyzed in relation to the predictors of interest which included: WM scores measured by the two tests; the amount of hearing loss represented by the average of pure-tone audiometric thresholds across 0.5, 1 and 2 kHz, averaged over both ears; and age (Age); by means of linear regression models. In the case of the AVDAT, only the four dual-modality components were included in regression analyses: the cued auditory selective attention task (DM-AC), the cued visual selective attention task (DM-VC), the auditory divided attention task (DM-AU), and the visual divided attention task (DM-VU). This was motivated by the fact that these components were the ones that tapped into divided and selective attention, while the other two components were simple digit/letter spans. Each of these four components was considered a separate predictor and entered separately in a regression model. The primary aims of the regression analyses were to assess: 1) the contribution of WM capacity in explaining additional variance in speech-in-noise recognition scores after the contribution of hearing loss (as measured by the PTA) was accounted for and 2) the individual contributions of WM capacity (as measured by RST and the AVDAT components) in explaining variance in speech-in-noise recognition (QuickSIN scores) regardless of hearing loss (i.e., as a stand-alone cognitive factor). Accordingly, multiple regression models were implemented in an incremental fashion for the first aim and simple regression models for the second. The multiple regression models included two predictors, wherein PTA was the first predictor, followed by one of the WM scores (RST or one of the AVDAT’s dual-modality components) or Age (the first column in Table 4 depicts the equations of the models that explained variance in QuickSIN scores). The WM scores from each test (RST and AVDAT) were entered in separate regression models (i.e., were never included in the same model) and in the case of the AVDAT, each component was entered in a separate model (i.e., no two or more AVDAT components were entered in the same model). All the numerical predictors were centered around their mean value before being included in the corresponding regression models. Each model was assessed for outliers after being fit with ordinary least squares linear regression (OLS). In the case of influential outliers (with a high Cook’s Distance) or with a large residual, robust regression (RR) was used for the model(s) in question to avoid any potential issues of problematic or over-estimation by the OLS model(s) in the presence of outliers (Cook, Hawkins, & Weisberg, 1992; Huber & Ronchetti, 2009). Residual and quantile plots of the linear models indicated that the assumptions of normality and linearity were satisfied (Hair, Black, Babin, & Anderson, 2010).

Table 4.

Model-comparison statistics for the effects of WM capacity (measured by the RST and the AVDAT) and hearing loss (represented by the PTA) on QuickSIN scores. The type of linear regression for each model is indicated in parentheses; OLS: Ordinary least squares regression; RR: Robust regression. For OLS models, the output of the model comparisons includes F, df and Pr(>F), and for the RR models the output consists of χ2, df and Pr(>χ2).

		Speech-in-Noise Recognition
PREDICTORS	Model	χ2/F	df	Pr(>χ2)/Pr(>F)
RST	M1 (OLS)	7.01	1	.02 (*)
PTA	M2 (RR)	9.46	1	.002 (**)
PTA + RST	M3 (RR)	5.61	1	.02 (*)
PTA + DM-AC	M4 (RR)	6.12	1	.01 (*)

Amount of hearing loss (PTA) is the across-ears average of hearing thresholds measured at 0.5, 1 and 2 kHz. RST: Reading Span Test; DM-AC: Dual Modality Auditory Cued AVDAT component.

The added effect of a predictor (improvement in the model fit) was assessed in an incremental fashion by performing likelihood ratio tests between the models with and without the predictor of interest. A predictor was included in a multiple regression model only if it contributed to a significant improvement in the fit of the model (explained additional residual variance).

As displayed in the correlation analysis in Table 3, RST and PTA were significantly correlated (r = −0.50, p = .03), however their variance inflation factors (VIFs) were < 2, indicating no serious concerns of multicollinearity that would impede their inclusion in the same linear model(s) (Hair et al., 2010). No significant correlations between PTA and any of the AVDAT components were found (Table 3). Table 4 displays the output of the model comparisons and Table 5 provides the summaries of the models with the most predictive power / best fit (simple and multiple regression models).

Table 5.

Summaries of the models wherein the addition of a predictor of interest, either alone (M₁ and M₂), or after the effects of hearing loss (PTA) had been controlled for (M₃ and M₄), explained significant variance in QuickSIN scores.

	Speech-in-Noise Recognition
Predictors
M1 (RST only)	β	SE	t	Pr(>|t|)	R 2 /Adjusted R 2
Intercept	3.87	.04	10.81	< .001 (***)
RST	−0.08	.03	−2.65	.02 (*)	.29 / .25
M2 (PTA only)	β	SE	t	Pr(>|t|)	R 2 /Adjusted R 2
Intercept	3.86	.35	11.06	< .001 (***)
PTA	.08	.03	3.08	.007 (**)	.32 / .28
M3 (PTA + RST)	β	SE	t	Pr(>|t|)	R 2 /Adjusted R 2
Intercept	3.80	.37	10.14	< .001 (***)
PTA	.06	.05	1.22	.24
RST	−0.05	.02	−2.37	.03 (*)	.43 / .36
M4 (PTA + DM-AC)	β	SE	t	Pr(>|t|)	R 2 /Adjusted R 2
Intercept	3.86	.31	12.52	< .001 (***)
PTA	.08	.03	3.04	.008 (**)
DM-AC	−1.02	.41	−2.47	.02 (*)	.46 / .39

Amount of hearing loss (PTA) is the across-ears average of hearing thresholds measured at 0.5, 1 and 2 kHz. RST: Reading Span Test; DM-AC: Dual Modality Auditory Cued AVDAT component.

As shown in Tables 4 and 5, participants’ performance on the reading span test (M₁) and their amount of hearing loss (M₂) were significant predictors of speech-in-noise scores (QuickSIN) when entered individually in the respective simple linear regression models. In line with our expectation, there was a main effect of verbal WM capacity measured by the reading span test (RST) on listeners’ sentence-in-noise scores, which is also consistent with existing literature on speech recognition in noise recognition performance. As displayed by the model summary in Table 5, listeners with higher WM capacities (higher RST scores) displayed better speech-in-noise scores. Importantly, as shown by the R² value (M₁), RST alone accounted for 29% of the variance in speech-in-noise performance (25% adjusted variance), comparable to the amount of variance explained by PTA alone (M₂), 32% (28% adjusted variance). None of the AVDAT components were significant predictors of speech-in-noise scores when included individually in the simple linear regression models.

RST remained a significant predictor of speech-in-noise scores after the effects of hearing loss were taken into account (M₃). The addition of RST as a predictor in M₃ resulted in an additional 8% of explained variance compared to the case where only the effect of hearing loss (PTA) was included in M₂ (adjusted R² difference between M₃ and M₂).

In regard to the AVDAT, only one of its components - the auditory cued dual modality task (DM-AC) – was a significant predictor after the effects of hearing loss had been controlled for (M₄). The inclusion of the dual-modality component with the cued auditory response (DM-AC) in M₄ led to an additional 11% of variance explained. There was no effect of age on QuickSIN scores after the effects of hearing loss were taken into account.

Discussion

Working memory, attention and speech recognition

The results of the present study revealed that performance on the RST was correlated to and predicted speech-in-noise recognition scores both before and after the effect of hearing loss were accounted for. This result was in line with our prediction and agrees with prior work that examined the relationship between WM capacity measured by the RST and speech-in-noise recognition (e.g., Souza & Arehart, 2015). The linear regression analysis revealed that on its own, the RST accounted for a comparable amount of explained variance in speech-in-noise scores to that accounted for by hearing loss alone. In addition, the RST remained a significant predictor of QuickSIN scores after the effects of hearing were accounted for in the statistical models.

In regard to the AVDAT, we found that one of its components which taps into selective attention, the cued auditory dual-modality task, was a significant predictor of speech-in-noise scores, in combination with hearing loss. This finding was in line with our prediction regarding selective attention and speech-in-noise recognition, and suggests that the task of recognizing speech in the presence of background noise relies on the ability to select in advance one source of information - the relevant/cued one – in the presence of irrelevant information competing for attention. This result extends those of Gallun and Jakien (2019), who found that the dual modality components of the AVDAT with the visual modality as the response (cued and uncued) were predictors of speech-on-speech performance. Contrary to our prediction regarding the relationship between divided attention and speech-in-noise recognition, none of the uncued dual modality components of the AVDAT that tap into divided attention were related to or predicted speech-in-noise scores. A possible explanation may rely on the different implementation of the divided/shared attention phenomenon in the two WM tests and on the extent to which the demands of the tests match with those of the speech task. In the case of the RST, the limited attention resource needs to be shared between the processing (semantic judgement of sentences) and storage (remembering the first and last words of the sentences) demands of the task. In the case of the ADVAT, there is no secondary task that requires processing of stimuli beyond merely attending to them and attention is shared between the two concurrent streams of non-word stimuli (digits and letters) that the participant needs to store for later retrieval. The speech-in-noise task requires the participant to recognize (i.e., process) the incoming words in the sentence stimuli and store them for integration with subsequent words and retrieval at the end of the sentence. As such, the sharing of attention between processing and storage of incoming speech input in the QuickSIN may display more overlap with the sharing of attention between processing and storage involved in the RST. Nevertheless, it should also be noted that while both the QuickSIN and the RST involve processing of words and sentence stimuli, the depth of processing may be different in each of them. That is, while the task demands of the RST require the participant to recognize and comprehend words in sentences in order to judge the sentence’s plausibility, the task of repeating back the words in a sentence in the QuickSIN may only require the recognition of the words, without evoking a deeper level of processing to comprehend them.

Overall, our results are in line with previous work demonstrating that WM capacity is related to recognition of speech in noise by hearing-impaired older listeners (e.g., Besser et al., 2013). In regard to the link between WM and attention, our results provide support for the storage and processing conceptualization of this link (shown via the RST) and more limited support for the controlled attention view (shown via one of the selective attention tasks of the AVDAT). While selective attention (shown via the cued auditory response dual modality component of the AVDAT) seemed to play a role in recognizing words in sentences degraded by the combination of background noise and hearing impairment, this role was limited to only one cued response modality (auditory). More research is needed to determine the role (if any) of the selective attention task of the AVDAT with the cued visual response modality in speech-in-noise recognition.

In contrast to Gallun and Jakien (2019), we found more limited evidence regarding the AVDAT and its relationship to speech recognition performance, with only one of its dual modality components being a significant predictor of speech-in-noise scores, compared to two dual modality components found to predict speech-in-speech performance in all the speech tasks of Gallun and Jakien (2019). Further, the components that were significant predictors of speech performance were different between the two studies: the cued auditory dual modality component in the present study and the cued and uncued visual dual modality components in Gallun and Jakien (2019). Methodological differences among our study and that of Gallun and Jakien (2019) may have been a factor in these differences. Specifically, our sentence stimuli were open-set, provide some semantic cues to the listener and display a certain degree of variability in their linguistic structure. The Gallun and Jakien (2019) sentence stimuli were closed-set, with an identical structure across them and with little-to-no linguistic and semantic information, which may have matched well with the visually cued and uncued dual-modality tasks of the AVDAT. Further, Gallun and Jakien (2019) measured speech-on-speech recognition in a multi-talker environment, which may evoke different phenomena from speech-in-noise recognition, such as informational masking, and consequently, employ different processing mechanisms.

To summarize, besides demonstrating the role of WM in speech recognition, our findings also indicate that a stronger match between the demands of the speech task and those of the WM test may capture the link between speech recognition and WM capacity more robustly compared to when the match between the demands of the speech task and the WM test is weaker. In comparison to the AVDAT, the combination of the cognitive demands and the type of stimuli of the RST seem to have been a better match for the cognitive demands and the type of stimuli of the speech-in-noise recognition task in the present study. Similarly, in Gallun and Jakien (2019), the demands and the type of stimuli of the speech-on-speech recognition tasks may have overlapped with the cognitive demands and the stimuli of the selective and divided attention tasks of the AVDAT to a larger extent compared to our speech recognition task and stimuli.

Which working memory test?

In the present study, the weak correlation between the two WM tests may indicate that they are not assessing the same construct to the same degree or manner. The more general and complex question of how to assess WM accurately and consistently across studies remains an open one. Our results suggest that a combination of factors that include methodological characteristics of the tests, the theoretical framework behind them and importantly, the match between the cognitive demands of the speech task and the cognitive abilities tapped by the WM tests, may govern the emergence of a relationship between WM and speech recognition. A relevant consideration for researchers studying the link between WM and speech recognition when choosing a WM test could be to decide what aspect(s) of WM and/or its relation to other cognitive abilities (such as attention) would be of interest for the speech recognition task in question. Lastly, although our sample size met the accepted range of suggested number of observations per predictor included in a regression model (Harrell, 2001), a larger sample would allow for replication and expansion to a wider range of hearing loss and/or participant age.

Data availability

The data that support the findings of this study are available from the corresponding author, [DS], upon reasonable request.