The association between cognition and speech-in-noise perception : Investigating the link between speech-in-noise perception and fluid intelligence in people with and without hearing loss

(1)

Linköping University | Department of Computer and Information Science Bachelor’s thesis, 18 hp | Cognitive Science Spring term 2020 | LIU-IDA/KOGVET-G--20/028--SE

The association between cognition

and speech-in-noise perception

- Investigating the link between speech-in-noise

perception and fluid intelligence in people with

and without hearing loss

Simon Dahlgren

Tutor: Erik Marsja Examinator: Rachel Ellis

(2)

Copyright

The publishers will keep this document online on the Internet – or its possible replacement – for a period of 25 years starting from the date of publication barring exceptional

circumstances.

The online availability of the document implies permanent permission for anyone to read, to download, or to print out single copies for his/her own use and to use it unchanged for non-commercial research and educational purpose. Subsequent transfers of copyright cannot revoke this permission. All other uses of the document are conditional upon the consent of the copyright owner. The publisher has taken technical and administrative measures to assure authenticity, security and accessibility.

According to intellectual property law the author has the right to be mentioned when his/her work is accessed as described above and to be protected against infringement.

For additional information about the Linköping University Electronic Press and its procedures for publication and for assurance of document integrity, please refer to its www home page: http://www.ep.liu.se/.

(3)

Abstract

The link between speech-in-noise recognition and cognition has been researched extensively over the years, and the purpose of this thesis was to add to this field. Data from a sample of 394 participants from the n200 database (Rönnberg et al., 2016) was used to calculate the correlation between their performance on a speech-in-noise test and their score on a test measuring fluid intelligence. The speech-in-noise test consisted of matrix sentences with 4-talker babble as noise and fluid intelligence was represented by the score on a Raven’s Progressive Matrices test. Around half of the participants (n = 199) had documented hearing loss and were hearing aid users, while the rest were participants with normal hearing. The overall correlation between speech-in-noise recognition and fluid intelligence was -.317, which shows that a better (lower) score on the speech-in-noise test is correlated to a better score in the Raven’s test. The same type of correlation was calculated within the two different groups, and the results showed correlation of -.338 for the group without hearing loss and one of -.303 for the group with hearing loss. The results indicate that there is a weak to moderate correlation between speech-in-noise and fluid intelligence, and they support the theory that cognitive processing is related to speech perception in all people, regardless of hearing status.

Keywords: cognition, speech-in-noise, fluid intelligence, Raven’s Progressive Matrices, hearing loss

(4)

List of figures

Figure 1: Correlation between 4-talker babble and score on Raven’s ... 13 Figure 2: Correlation for the group without hearing loss. ... 14 Figure 3: Correlation for the group with hearing loss. ... 15

(6)

1

1. Introduction

The ability to communicate with others is essential for all people. One of the most important abilities in communication is listening; to hear and understand what others are trying to say. Work in the field of cognitive hearing science has shown that this is not only related to the ability to receive input with our sense of hearing, but also how the input is processed using different cognitive functions (Rönnberg, Rudner & Lunner, 2011). Another important factor in verbal communication is the amount of surrounding noise that may interfere with the ability to recognise what others are saying, which is why research on speech-in-noise recognition is useful in understanding the mechanisms that people use to understand each other in everyday life. Although most of the evidence on cognitive processing in language understanding has been found in participants with hearing loss, there is evidence that suggests that this happens in all people in challenging listening conditions (Rönnberg, Holmer & Rudner, 2019).

The link between cognition and speech-in-noise perception has been researched extensively. Dryden, Allen, Henshaw, and Heinrich (2017) assessed the field in a literature review and meta-analysis and found that there is a positive correlation between general cognitive performance and speech-in-noise recognition performance. There is, however, more research to be done in this field. Cognition is generally considered to encompass a range of different functions, for example memory, executive processes, and intelligence among others. Some of these domains can be further separated into subcategories, e.g. working memory and episodic memory (Dryden et al., 2017). To fully understand the relation between cognition and speech-in-noise perception, all different aspects of human cognition should be studied.

In the work of Dryden et al. (2017) five main cognitive domains and nine subdomains were defined. While some of these areas, for example working memory, were found to have a stronger correlation to speech-in-noise performance than others, some of them lacked enough data for a meta-analysis to be made. Much like how memory is studied in subcategories, the domain intelligence is frequently divided into crystallized intelligence and fluid intelligence in psychological research. While Dryden et al. did have enough data to analyse the role that crystallized intelligence plays in relation to speech-in-noise in their review, they described fluid intelligence as a relatively underinvestigated area where further research would be valuable, as it has an established link to working memory as well.

(7)

Since both speech-in-noise recognition and fluid intelligence are related to working memory, and fluid intelligence is one of the least studied cognitive domains when it comes to speech-in-noise recognition and cognition, the purpose of this study is to investigate the link between fluid intelligence and speech-in-noise recognition. Since a large part of the research within this field has been done on participants with hearing loss, this study will also investigate the link between these variables in a group of participants with hearing loss as well as a group of participants with normal hearing.

1.1. Research questions

The research questions for this study are:

1. Is speech-in-noise recognition related to fluid intelligence?

2. Is there a difference in the significance of the correlation between speech-in-noise recognition and fluid intelligence depending on if the participant has documented hearing loss or not?

(8)

2. Theory

2.1. Speech-in-noise recognition

Speech-in-noise tests are a valuable tool in research of how people recognise and process verbal communication in real-life situations. It has been used extensively in research, for example in research on the rehabilitation of people with hearing loss (Rönnberg, Holmer & Rudner, 2019). Because of this, a large part of the research within the field of cognitive hearing science, which studies the connection between hearing and cognition, has been conducted on participants with hearing loss. Hearing loss is a reliable predictor of speech reception (Akeroyd, 2008), but Dryden et al. (2017) explain that it cannot be the only factor that is responsible for speech-in-noise performance, as people without any hearing impairment can have trouble with it as well. In addition to this, people with the same auditory sensitivity can perform differently in these tests, and introducing a hearing aid does not always remove people’s difficulties. This might be because cognitive factors affect speech recognition; Dryden et al. (2017) report a general correlation of r ≈ .3 between speech perception and cognitive performance.

There are different kinds of speech-in-noise tests, as different kinds of target speech (for example words or sentences) and different kinds of background noise can be used. In Akeroyd’s review (2008) it was found that studies that used sentences as target speech found a relation between speech-in-noise and cognition to a bigger extent than studies that used words, for example. The conclusion that was drawn from this is that the higher complexity of sentences and the fact that sentences require more resources to recognise than single words leads to a stronger effect with respect to cognition.

One type of sentence that is commonly used in speech-in-noise tests is matrix sentences, for example the Hagerman sentences (Hagerman, 1982; Hagerman & Kinnefors, 1995). Matrix sentences are based on a simple and predictable grammar. In the case of the Hagerman material, all the sentences consist of a proper noun, a verb, a numeral, an adjective, and a noun, in that particular order. While the grammatical structure of the sentences is predictable, the words themselves are not predictable, and they cannot easily be guessed if they are not heard (Rönnberg et al., 2019). This is in contrast to other types of sentences, for example the Hearing in Noise Test (HINT) sentences (Hällgren, Larsby & Arlinger, 2006). In HINT tests, the sentences are naturalistic and have a context which makes speech recognition easier, because it provides a better opportunity to anticipate what is going to be said at the end of a sentence after hearing the beginning of it.

(9)

When it comes to background noise, Dryden’s et al. (2017) review divided it into three general types: unmodulated noise, modulated noise, and babble, where babble was divided into >2-talker babble, meaning that there were at least three background talkers used as noise; and ≤2-talker babble, where only one or two background talkers were used. The babble conditions had more informational masking than the others, because the noise had sounds that were intelligible for the participants.

2.2. Working memory and speech-in-noise

Defining the concept of working memory is not the easiest task, as it has been used in different ways in different fields. In a well-cited article, Baddeley (2000) explains that working memory is a system with limited capacity, where information is stored temporarily. The stored information is used for complex tasks such as reasoning, learning, and comprehension. A person’s working memory capacity is their ability to store and process information simultaneously (Rönnberg et al., 2013). A link between working memory and speech understanding has been found in several studies, and a large part of the literature found this connection in people with hearing loss (e.g. Lunner, 2003; Foo, Rudner, Rönnberg & Ludner, 2007; Arehart, Souza, Baca & Kates, 2013). There is however evidence that suggests that people with normal hearing use explicit cognitive processing in challenging listening conditions as well, as Michalek, Ash, and Schwarz (2018) found that more challenging signal-to-noise ratios in babble noise increased participants’ dependence on working memory capacity. In a review, Akeroyd (2008) found that although hearing loss was a big predictor of speech recognition in noise, working memory was mostly effective in predicting performance as well. This shows that both a person’s hearing and their cognitive status affect how well they can recognise speech in noise, to some degree.

One prominent theory when it comes to the role that working memory plays in speech recognition is the Ease of Language Understanding (ELU) model (Rönnberg, 2003; Rönnberg et al., 2019). It is a model that describes the mechanisms that affect language understanding in different everyday listening conditions. Some of these conditions may push the cognitive processing system to a point where the ease of language understanding is reduced, for example if the speech is masked by background noise (Rönnberg et al., 2019). There are several other factors that are relevant to ELU, including familiarity of language and accent, speed and flow of conversation, and how coherent the speech is when it comes to phonology, grammar, and

(10)

semantics. If these factors are not optimal, for example if there is a large amount of background noise, it leads to challenging listening conditions.

The ELU model is a framework that highlights the role that cognitive processing plays in language understanding (Rönnberg et al., 2019). This means that it is closely related to the field of cognitive hearing science, which has emerged in the last two decades, and focuses on the interactions between hearing and cognition in humans (Arlinger, Lunner, Lyxell, & Pichora-Fuller, 2009). According to ELU, speech input is firstly processed in the so-called RAMBPHO (rapid, automatic, multimodal binding of phonology) buffer, which binds representations based on the multimodal information relating to syllables. If the input matches a phonological and lexical representation in the listener’s semantic long-term memory, lexical access is successful, and the listener recognises what has been said.

If, however, there is a mismatch between the input and the representations in the listener’s long-term memory the RAMBPHO buffer feeds it forward to working memory and executive functions (Rönnberg et al., 2019). Rönnberg proposes that this triggers a slower inference-making process using working memory, in order to attempt to reconstruct what was not recognised. These mismatches can occur as a result of a number of different factors, for example speech rate and hearing status. Additionally, if the mismatch is too great, the listener might not be able to recognise the speech through this process.

In the ELU model, working memory capacity has two roles when it comes to language understanding: prediction and postdiction (Rönnberg et al., 2013, 2019). The postdictive role refers to the cognitive mechanism that comes into play when a phonological mismatch has occurred, as mentioned earlier. This is where working memory is invoked for some kind of inference-making regarding what was heard. The predictive role of working memory, according to the model, concerns priming and pre-tuning the RAMBPHO buffer and focussing of attention. Rönnberg et al. (2013) explain that phonological information as well as semantic information is used for prediction, and there is evidence that suggests that using this predictive ability to perceive speech is related to working memory capacity when the speech input is degraded (Hunter & Pisoni, 2018; Zekveld, Kramer, and Festen, 2011; Zekveld, Rudner, Johnsrude, Heslenfeld, & Rönnberg, 2012). The predictive function of working memory is assumed to work quickly, on a scale of tenths of seconds, while the postdictive function is assumed to be relatively slow and deliberate, operating on a scale of seconds (Stenfelt & Rönnberg, 2009).

The ELU model explains the role that working memory may have in speech-in-noise recognition, but there is also a difference in how important working memory capacity is

(11)

depending on the type of background noise and target speech used. In Ng and Rönnberg (2020), 4-talker babble was compared to stationary noise, and the results showed that the participants’ dependence on working memory capacity was significantly higher in the 4-talker babble condition than for stationary noise. Ng and Rönnberg used the n200 database (see the method section of this thesis for details on the database), which means that they had a large sample size. It should be noted that these findings were on people with hearing loss, however.

When it comes to different types of target speech, Rönnberg et al. (2016, 2019) explain that matrix sentences (like the Hagerman sentences) increase participants’ dependence on working memory, while sentences with a higher degree of contextual support (like the HINT sentences) reduce the dependence on working memory. This is because the higher lexical predictability and clearer context of the latter reduce the need for cognitive functions to make inferences about what was said.

All research in the field of cognitive hearing science does not support the ELU model. In fact, Füllgrabe and Rosen (2016) claim that we should be cautious in assuming that the ELU model explains how all people process speech in noise. They reviewed 19 studies that measured working memory and speech-in-noise in people with normal hearing, and performed a meta-analysis. They found that, on average, working memory capacity accounts for less than 2% of the variance in speech-in-noise scores in young participants with audiometrically normal hearing. Füllgrabe and Rosen explain that there might be a tendency for researchers to assume that working memory capacity plays the same role in speech-in-noise recognition in all people, even though most of the evidence is found in older people and people with hearing loss. Their study shows that this assumption cannot be made without providing further evidence for this. As previously mentioned, there is some evidence that indicates that a higher working memory capacity is related to better speech-in-noise recognition in participants without hearing loss (Michalek et al., 2018), but as Füllgrabe and Rosen (2016) found their results through a meta-analysis of 19 different studies, their conclusions should be taken into account as well.

2.3. Fluid intelligence

As previously mentioned, intelligence is frequently divided into crystallized and fluid intelligence in psychology. Crystallized intelligence refers to language- or culture-related knowledge and skills which are acquired over time. It is consequently dependent on the person’s cultural and educational background (Dryden et al., 2017; Engle, Tuholski, Laughlin &

(12)

Conway, 1999). Fluid intelligence, on the other hand, refers to the general ability to solve novel problems using abstract reasoning. It allows us to recognise patterns and engage with the world, and it has been shown that individual differences in fluid intelligence can be related to academic and occupational success (Cochrane, Simmering & Green, 2019). Kyllonen and Christal (1990) describe fluid intelligence as being close to what is ordinarily meant when people use the word “intelligence”. They also point out that fluid intelligence (also known as reasoning ability) may consist of three types of reasoning: deductive reasoning, inductive reasoning, and quantitative reasoning. These three factors are hard to separate, however, as tasks used to test one of the reasoning abilities often involve the other two as well.

Fluid intelligence has been shown to be strongly linked to working memory capacity. Kyllonen and Christal (1990) performed four studies which aimed to determine what the relationship is between the two factors. The four studies had over 2000 participants collectively, and a variety of tests were used to measure fluid intelligence and working memory capacity. The working memory tests were created using Baddeley’s (1986) definition of working memory as a base. In all of their studies, they found a high correlation between the two factors (.82, .88, .80, and .82 in studies 1 through 4, respectively), which they considered interesting, as the two concepts had formed from two different parts of psychology (Kyllonen & Christal, 1990).

In a different study, Prabhakaran, Smith, Desmond, Glover, and Gabrieli (1997) used fMRI to study brain activity while the participants performed the Raven’s Progressive Matrices test, which is a commonly used test to measure fluid intelligence. They found a strong link between working memory and results on the Raven’s tests and proposed that the reason might be that the tasks used to measure both of the processes are measuring common neural systems. They also found that “performance on the Raven’s Progressive Matrices reflects the status of numerous, perhaps almost all, working memory systems” (Prabhakaran et al., 1997, p. 60).

In a more recent study, Cochrane et al. (2019) explain that the relation between working memory and fluid intelligence is well-established and has been studied in depth. The purpose of their study was to see whether performance on other psychological tasks was related to these two cognitive functions. They had children and adults perform cognitive tasks, including working memory and reasoning tasks. In accordance to previous studies, they did find a strong link between working memory and fluid intelligence, but the other main domain that correlated with fluid intelligence was attention. Further analysis showed that working memory fully mediated the link between attention and fluid intelligence, while attention did not mediate the link between the other two.

(13)

2.4. Raven’s Progressive Matrices

As was mentioned in the last section of this thesis, Raven’s Progressive Matrices test is a commonly used test to measure general fluid intelligence. Cochrane et al. (2019) describe it as the gold standard for reasoning measures, which is used in many studies on the cognitive bases of fluid intelligence, as well as in cognitive training paradigms. Raven’s Progressive Matrices is a nonverbal test created to measure what researchers often call “general cognitive ability” (Raven, 2008). Raven (2008) explains that it measures a kind of meaning-making ability, to construct meaning out of something abstract and confusing. In the test, participants are presented with figures in certain patterns, generally a grid of nine figures where the ninth figure is missing. For each item there are six alternatives, and participants are asked to choose which alternative completes the pattern in question.

There are several different versions of the Raven test, and two of the most well-used ones are the standard version and the advanced version (Raven 2008). The Standard Progressive Matrices test is the most basic Raven’s test, and it covers all levels of ability. It does not even require the ability to read, which means that is can be used on all participants, from early childhood to old age, and on people from different socio-economic backgrounds. Raven’s tests also have a low cultural load, since all the items consist of geometric figures (Arthur & Day, 1994). This means that no special cultural knowledge is needed to participate, and they can be used in many different cultures. The original test consisted of 60 problems, divided into five sets of 12 problems each (sets A, B, C, D, and E) (Raven, 2008). Within each set the items become progressively more difficult, but at the start of the next set the difficulty reverts back to an easier level. Raven (2008) points out that theoretically, participants should answer correctly on every question until their limit and get the rest of the questions above that limit wrong.

The Advanced Progressive Matrices test is a different version of Raven’s, and it is used extensively in research on adults with particularly high ability (Arthur & Day, 1994). This version has harder items than the original version. Jensen, Saccuzzo, and Larson (1988) compared the advanced version of the test with the standard version, as they wanted to compare results from two different groups. They recruited participants from U.C. Berkeley, which they described as a university that is highly selective when it comes to which students they accept. In addition to this, they recruited another group of participants from a community college, where the only entrance requirement was a high school diploma. In their pilot study, Jensen et al. (1988) found that none of the test versions worked perfectly for both participant groups; the

(14)

standard test was too easy for the highly capable university students, while the advanced test was too difficult for the community college students. Some of the community college students performed no better than the chance prediction for such a multiple-choice test. With this in mind, it would seem that the Standard Progressive Matrices test would fit better in studies where the participants are diverse in terms of cognitive performance, while the advanced test fits better in studies where the participants are known to be from a higher-ability population.

Due to the high number of items in the Raven’s Standard Progressive Matrices, and the complexity of them, one potential problem is the amount of time that is needed to finish the full test. If the experimenter needs to save time on the test administration, one option is to introduce a time limit (Bilker et al. 2012). The Raven’s tests are meant to be untimed, however (Raven, 2008), and if a time limit is introduced it can be argued that the test is no longer testing the exact thing it is supposed to test (Bilker et al., 2012). Because of this, there have been studies made on how the test can be abbreviated in a way that reduces the time needed to complete it, but still measures fluid intelligence accurately. An example of this is Arthur and Day (1994) who reduced the number of items in the Advanced Progressive Matrices test from 36 to 12, while retaining similar psychometric properties to the original test.

Another example of significantly reducing the number of items in a Raven test can be found in the study by Bilker et al. (2012). They used an algorithm to create two new versions of the Standard Progressive Matrices, with only nine items from the original test in each new test. Since the original test has 60 items, this is a reduction of 85% in the number of items used, which leads to a large decrease in the amount of time needed to administer the test. In addition to this, the two 9-item tests that Bilker et al. created had strong correlations to the long form, and the psychometric properties were found to be comparable to those of the original Standard Progressive Matrices as well. This shows that an abbreviated version of the Raven’s test may provide an accurate and useful result.

(15)

3. Method

3.1. Participants

The data that were used in this study were acquired from the n200 database, which was created by Rönnberg et al. (2016). The database consists of three participant groups with around 200 participants in each group. For this study, participants from two of these groups were used in the analyses; one of the groups were people with documented hearing loss who were used to wearing a hearing aid, while the other group consisted of people without any hearing loss, who consequently were not hearing aid users. The number of participants in the dataset used in the current study was 433 (207 females, 221 males, and 5 participants whose gender information was missing). 215 of these were hearing aid users with hearing loss, and the remaining 218 did not have hearing loss. The age of the participants ranged between 27-81 years (M = 61.1 SD = 8.58) As some (or in a few cases almost all) of the information was missing from a small number of participants, the number of participants used in the analyses is 394.

In the recruitment procedure, the participants were sent an information letter which stated that participation was voluntary and could be ended at any time (Rönnberg et al. 2016). It was also made clear in the letter that the participants would remain anonymous and that their data would be stored in a database. The testing was divided into three separate sessions, which lasted around 2-3 hours for every participant. The first session included hearing tests, the second session consisted of the cognitive tests, and in the third session the participants’ speech recognition in noise was tested. Rönnberg et al. (2016) explain that the stimuli in all the tasks in the database were balanced for relevant parameters when needed, and that the order of the trials within the tasks was randomized and fixed for all participants before the start of the experiment. As the n200 database contains a large battery of different tests, only the ones that were used in the analyses of this thesis will be described in further detail below.

3.2. Raven’s Progressive Matrices

A shortened version of the Raven’s Standard Progressive Matrices test was used to measure fluid intelligence (Rönnberg et al., 2016). Three of the five test sets, specifically sets A, D, and E, were completed by the participant using pen and paper. Each set consisted of 12 items with increasing difficulty, and the participant marked for each item which of the six available alternatives they thought was the missing piece of that particular item’s pattern. At the start of

(16)

the test set A was administered for practice, and the experimenter was allowed to give feedback to the participant at this point. After this, the experimenter administered the two remaining sets (D and E) for the participant to complete without getting any feedback or time limit. The test was scored based on the participant’s sum of points on test sets D and E, where the maximum score for each set was 12 points, and the maximum score for the entire test was consequently 24 points.

3.3. Hagerman speech-in-noise

The n200 database contains results from different types of speech-in-noise tests, both with regard to sentence type and type of background noise (Rönnberg, et al., 2016). The test that was used in the analyses of this current thesis was the Hagerman test with 4-talker babble background noise. In this speech-in-noise test, the participant listened to sentences that consisted of a proper noun, a verb, a numeral, an adjective, and a noun, in that order (Example: “Ann had five red boxes”) with background noise consisting of four talking people. After hearing each sentence, the participant repeated the sentence that they heard verbally. The experimenters used an adaptive procedure to estimate the signal-to-noise ratios (SNR) where the participant could repeat 50% and 80% of the words correctly.

The participants were tested with a type of experimental hearing aid in an anechoic chamber through a measuring amplifier (Rönnberg et al. 2016). Two types of amplification were used:

linear amplification and non-linear fast-acting compression. A noise reduction algorithm (the

binary mask NR algorithm) was also used, which reduced the masking effect of the interfering speech noise. This is done by removing noise-dominant spectro-temporal regions of the speech-in-noise material. The technical details of these systems are explained further in the supplementary material of the original n200 article of Rönnberg et al. (2016). The use of the systems resulted in three different test conditions used in this present study:

1. Linear amplification without noise reduction in 4-talker babble noise.

2. Linear amplification combined with noise reduction in 4-talker babble noise.

3. Fast-acting compression signal processing in 4-talker babble noise (without noise reduction).

In each of the conditions, the experimenters recorded the SNRs (measured in dB) where the participant could correctly repeat 50% of the words that were heard, and also the SNR at which 80% of the words were correctly repeated (Rönnberg et al., 2016). The result of this is that each test condition essentially worked as two conditions, as two different SNR levels were recorded

(17)

for every participant. As the SNR measurements represent the strength of the target signal in relation to the surrounding noise, the lower this number is the harder it is to recognise what is said. Because of this, a lower number on this speech-in-noise test is what would be regarded as a better score.

3.4. Analysis

For the analyses in this study, the mean of the six SNR variables from the Hagerman speech-in-noise test was calculated for each participant, forming a new single Hagerman 4-talker babble variable. To examine this variable’s relation to fluid intelligence, correlation coefficients were calculated between it and the participants’ scores on the Raven’s Progressive Matrices test. In order to inspect whether there were any differences in the significance of the correlation depending on hearing loss, the correlations were calculated within the two participant groups as well.

Before the statistical analyses, the data were examined in the form of histograms, box plots, and Q-Q plots to determine whether the assumptions for parametric tests were fulfilled. The Raven variable appeared to break the assumption of normality, and outliers were found in the 4-talker babble variable. While normality might be less of a concern when the sample size is large, this assumption violation in conjunction with the presence of outliers in the data led to the utilisation of nonparametric tests.

(18)

4. Results

In this section, the results of the statistical analyses are presented.

4.1. Speech-in-noise and fluid intelligence for all participants

Firstly, the correlation between the mean of the six Hagerman variables and the participants’ scores on the Raven’s test was calculated. In this correlation there were 394 participants. Spearman’s rank correlation coefficient showed a significant negative correlation, rs = -.317, p

< .001. This means that recognising speech at a lower SNR is related to a higher score on the intelligence test. This correlation is represented in Figure 1 below.

Figure 1: Scatterplot illustrating the correlation between the combined 4-talker babble variable and

score on Raven’s Progressive Matrices

4.2. Participants without hearing loss

Next, the Spearman correlation coefficient was calculated between the Hagerman mean score for the group without hearing loss and the Raven’s Progressive Matrices scores for the same

(19)

group. Here, there were 195 participants. The correlation found in this group was similar to the correlation in the first analysis: rs = -.338, p < .001. Figure 2 shows this correlation in a

scatterplot.

Figure 2: The correlation between speech recognition in noise and Raven’s test scores for the

group without hearing loss.

4.3. Participants with documented hearing loss

Lastly, the Spearman correlation was conducted for the participant group who had documented hearing loss and was experienced with hearing aids. The number of participants in this analysis was 199. Their speech-in-noise scores also showed a similar negative correlation to their Raven’s test scores: rs = -.303, p < .001. This correlation is represented in Figure 3. Note that

the statistics software used to generate the scatterplots automatically used a slightly different scale of values on the y-axis in this figure when compared to the others; showing values between -2,5 and 2,5 instead of between -5 and 5.

(20)

Figure 3: The correlation between speech recognition in noise and Raven’s test scores for the

(21)

5. Discussion

5.1. Results discussion

In order to advance the field of research on the relation between speech-in-noise recognition and cognition, the purpose of this study was to investigate whether speech recognition in noise and fluid intelligence were related, and whether the significance of the correlation would look different in people with documented hearing loss when compared to people with normal hearing. The previous research reviewed in this thesis shows that both speech-in-noise recognition and fluid intelligence have an established connection to working memory, and Dryden’s et al. (2017) review of the field showed that fluid intelligence was one of the least researched areas of cognition when it comes to the relation to speech-in-noise recognition. These findings gave rise to the research questions comparing the two variables directly. To answer the research questions, data from the n200 database (created by Rönnberg et al., 2016) was used, resulting in approximately 400 participants. For speech-in-noise, a test with matrix sentences and multi-talker babble was used, as previous research had showed that these conditions lead to an increased dependence on working memory, which, just like fluid intelligence, is regarded as a part of human cognition. Fluid intelligence was measured by a shortened version of the commonly used Raven’s Standard Progressive Matrices test.

In order to assess the correlation between the relevant variables, Spearman’s rank correlation coefficient was calculated. The speech-in-noise variable (which was a combination of six variables from the three different hearing aid conditions described in the Method section of this thesis) had a significant negative correlation of -.317 to score on the Raven’s Progressive Matrices test. As previously mentioned, speech-in-noise performance was measured as signal-to-noise ratio (SNR) in dB, where the participants could recognise and repeat 50% and 80% of the words heard, respectively. This means that a lower number in this variable indicates a weaker target signal in relation to the noise signal, which is a better performance on the test. With this in mind, the negative correlation shows that a better speech-in-noise score is related to a better score on fluid intelligence.

The dataset used consisted of two participant groups, one of which was people with hearing loss who were used to using a hearing aid. The other group consisted of people with normal hearing. Spearman’s rank correlation coefficient was calculated between speech-in-noise recognition and Raven’s test score within these groups as well, showing a correlation of -.303 for the participants with hearing loss, and a correlation of -.338 in the group without hearing

(22)

loss. Both of these were statistically significant. As the correlation for the group with hearing loss was slightly weaker, they could possibly have more dependence on the hearing aid while the other group used cognition to a higher degree. However, since the difference between the two correlations is small, and the difference between the correlations has not been checked statistically, this argument is not particularly strong, and would need further research.

In general, this study generated interesting results. As the three correlations found in this study are a little stronger than -.3, speech-in-noise recognition and fluid intelligence are related to each other, but only to a weak to moderate degree. If one reviews the rest of the field, however, it can be seen that the strength of correlations between speech-in-noise and cognitive factors is often around .3 (see the review written by Dryden et al. (2017) for example), which means that the results from this study fit and work well in conjunction with previous findings. When it comes to hearing loss, this study found that the size of the correlation for people with hearing loss was very similar to the correlation for people with normal hearing. As helping people with hearing loss has been an important focal point for cognitive hearing science, a large part of the conducted research has used participants with these kinds of impairments. As a result, some findings within the field cannot necessarily be generalised to people without hearing loss. The results from this study can be interpreted as supporting the theory that all people use cognitive processing when listening to speech in noise, regardless of hearing status. This can be interpreted as supporting the ELU model (Rönnberg 2003; Rönnberg et al., 2013, 2019), where cognitive processing comes into play in challenging listening conditions in all people.

Speech-in-noise tests are used to simulate everyday listening conditions where there are other sounds competing with the speech one is trying to hear. The findings of this study indicate that the ability to successfully recognise speech in noisy environments is stronger in people with a higher fluid intelligence, which means that they are better at using abstract reasoning to solve problems and see patterns. This is the case regardless of the hearing status of the listener. It could be that this abstract reasoning plays a role in the postdictive process described in the ELU model, in order to make inferences about what was said. The findings are in any case an interesting addition to the field of cognitive hearing science, as they provide insight into the relation between speech-in-noise recognition and cognition.

5.2. Method discussion and future studies

(23)

A major strength of this study is its statistical power. Using the data from the n200 database resulted in over 400 participants, 394 of which could be used in the relevant analyses. Such a large sample size provides an excellent opportunity to perform analyses without needing to worry about the difficulties that come with small sample sizes (e.g. overestimated effect sizes and low reproducibility of results (Button et al., 2013)), and the statistical power increases the probability that the effect will be found, if there is one.

In the n200 database, there are several different speech-in-noise tests. The choice of matrix sentences and 4-talker babble was based on previous research indicating a higher dependence on cognitive functions in these conditions, as mentioned before (Ng & Rönnberg, 2020; Rönnberg et al., 2016, 2019). In the analyses in this current study, there were however other choices made that affected the results that could be generated. The six different Hagerman test variables (from three hearing aid conditions, each with a variable where 80% of the words were correctly repeated and one where 50% of the words were correctly repeated) were combined into one variable which calculated the mean of the six variables for every participant. This was done in order to get a view of general speech-in-noise performance, but in future studies it could be of interest to investigate the different conditions separately. There could for example be a difference between linear amplification and fast-acting compression when it comes to the correlation to fluid intelligence, or other cognitive factors. Another related aspect that could be investigated in future studies is whether the relation between fluid intelligence and speech-in-noise recognition is different when the participant can correctly repeat 50% of words, in comparison to the harder task of repeating 80% of the words. Perhaps the reliance on cognition is higher in one of these cases.

Future research can explore the relation between speech-in-noise and cognitive factors further. As working memory capacity and fluid intelligence are often found to be highly correlated, it would be interesting to examine whether fluid intelligence plays a role in the effect between working memory and speech-in-noise. Since a correlation was found between speech recognition in noise and fluid intelligence in this current study, and this correlation is of similar strength to correlations between cognition and speech-in-noise that have been found in previous studies (e.g. Dryden et al., 2017), future studies could for example look at the established relation between cognition and speech-in-noise, while controlling for fluid intelligence or other cognitive factors. Perhaps fluid intelligence is the cause behind some of the effects that have previously been attributed to working memory or other cognitive domains.

(24)

memory capacity, there might be a difference in the dependence on cognitive factors in speech-in-noise recognition when comparing different age groups. Future studies could investigate whether the link between fluid intelligence and speech recognition in noise is different in older participants when compared to younger participants.

This study used participants’ scores on Raven’s Standard Progressive Matrices as a representation of their fluid intelligence. This can be regarded as a strength, as Raven’s tests are well-used and popular in fluid intelligence research. The fact that they cover a wide range of ability, and that finishing the test does not require much prior knowledge or cultural experience makes it useful in a wide range of situations. One potential weakness could be that previous research (e.g. Jensen et al., 1988) has shown that the items in the standard test can be too easy for the ones with the highest ability. While the advanced version of the test could have been better to measure the ability of these participants, this would have made the test too hard for the other participants, resulting in a large number of low scores. The mean age of the participants in this current study is also relatively high, which indicates that they might not have the same cognitive capacity as the college students in the study of Jensen et al. (1988) With this in mind, Rönnberg et al. (2016) likely made the right choice in choosing the standard version of the Raven’s test for the n200 database, where a large number of participants were tested.

In the database, the version of the Raven’s Standard Progressive Matrices test used was a shortened version of the original test. This was done by choosing three of the five available item sets, using one of the sets for practise and using the remaining two for the actual test. This reduced the number of items administered from 60 to 24. As described in section 2.4 of this thesis, the Raven’s tests can be shortened without sacrificing much of their psychometric properties. While Bilker et al. (2012) had a more advanced method of shortening the test, their shortened versions of the Raven’s Standard Progressive Matrices consisted of only nine items each, while retaining the properties of the original to a large extent. While Rönnberg et al. (2016) used a simpler method to make the test shorter for the n200 database, the number of items in their test is significantly higher at 24. The validity of their measurement of fluid intelligence is therefore likely high, but it should be kept in mind that it might not be perfect.

In this study, the relation between speech-in-noise recognition and fluid intelligence was investigated using correlation coefficients. This is a relatively simple method of analysis, and it has the inherent issue that nothing can be said about the causality between the variables in question. When looking at the research field as a whole, however, it is shown that correlations are a frequently used method of analysis when looking at speech recognition in noise and cognition (see the reviews of Akeroyd (2008) and Dryden et al. (2017)). Other studies could

(25)

however apply other types of statistical tests and do different kinds of group comparisons than those that have been done in this current study.

(26)

6. Conclusion

The purpose of this study was to investigate the relation between speech-in-noise recognition and fluid intelligence. The results show a significant correlation of -.317 between scores on the Hagerman 4-talker babble speech-in-noise test and higher scores on the fluid intelligence test Raven’s Progressive Matrices. As a lower score in the speech-in-noise test is regarded as better, a better speech-in-noise recognition is correlated with a higher fluid intelligence. Computing the same correlations in a group of experienced hearing aid users with hearing loss and a group of people with normal hearing shows that there is a very similar relationship between speech-in-noise and fluid intelligence regardless of hearing loss. The large sample size of this study, as well as the fact that the correlations found were of a similar size to correlations found in previous studies looking at cognitive functions in relation to speech-in-noise, adds to the credibility of the results.

(27)

7. Bibliography

Akeroyd, M. A. (2008). Are individual differences in speech reception related to individual differences in cognitive ability? A survey of twenty experimental studies with normal and hearing-impaired adults, International Journal of Audiology, 47,

53-71, DOI: 10.1080/14992020802301142

Arehart, K. H., Souza, P., Baca, R., & Kates, J. M. (2013). “Working Memory, Age, and Hearing Loss: Susceptibility to Hearing Aid Distortion.” Ear and Hearing, 34(3): 251– 260. doi:10.1097/AUD.0b013e318271aa5e.

Arlinger, S., Lunner, T., Lyxell, B. & Pichora-Fuller, M. K. (2009). The emergence of Cognitive Hearing Science. Scandinavian Journal of Psychology, 50, 371–384.

Arthur, W., & Day, D. V. (1994). Development of a Short form for the Raven Advanced Progressive Matrices Test. Educational and Psychological Measurement, 54(2), 394– 403. https://doi.org/10.1177/0013164494054002013

Baddeley, A.D. (1986). Working memory. Oxford: Clarendon Press.

Baddeley A. (2000) The episodic buffer: a new component of working memory? Trends in

Cognitive Sciences, 4(11), 417–423.

Bilker, W. B., Hansen, J. A., Brensinger, C. M., Richard, J., Gur, R. E., Gur, R. C. (2012). Development of Abbreviated Nine-Item Forms of the Raven’s Standard Progressive Matrices Test. Assessment, 19(3), 354-369. DOI: 10.1177/1073191112446655

Button, K. S., Ioannidis, J. P. A., Mokrysz, C., Nosek, B. A., Flint, J., Robinson, E. S. J., & Munafò, M. R. (2013). Power failure: Why small sample size undermines the reliability of neuroscience. Nature Reviews Neuroscience, 14(5), 365-376. https://doi.org/10.1038/nrn3475

(28)

Cochrane A, Simmering V, Green CS (2019) Fluid intelligence is related to capacity in memory as well as attention: Evidence from middle childhood and adulthood. PLoS ONE 14(8).

https://doi.org/10.1371/journal.pone.0221353

Dryden, A., Allen, H. A., Henshaw, H., & Heinrich, A. (2017). The Association Between Cognitive Performance and Speech-in-Noise Perception for Adult Listeners: A Systematic Literature Review and Meta-Analysis. Trends in Hearing, 21, 1–21.

Engle R. W., Tuholski S. W., Laughlin J., Conway A. R. A. (1999). Working memory, short-term memory and general fluid intelligence: A latent variable model approach. J Exp

Psychol Gen. 128(3) 309–331. https://doi.org/10.1037/0096-3445.128.3.309

Foo, C., Rudner, M., Rönnberg, J., & Lunner, T. (2007). “Recognition of Speech in Noise with New Hearing Instrument Compression Release Settings Requires Explicit Cognitive Storage and Processing Capacity.” Journal of the American Academy of

Audiology, 18, 553–566.

Füllgrabe, C., & Rosen, S. (2016). On The (Un)importance of Working Memory in Speech-in-Noise Processing for Listeners with Normal Hearing Thresholds. Frontiers in

psychology, 7, 1268. https://doi.org/10.3389/fpsyg.2016.01268

Hagerman, B. (1982) Sentences for Testing Speech Intelligibility in Noise, Scandinavian

Audiology, 11(2), 79-87. https://doi.org/10.3109/01050398209076203

Hagerman, B., Kinnefors, C. (1995). Efficient Adaptive Methods for Measuring Speech Reception Threshold in Quiet and in Noise. Scandinavian audiology, 24, 71-7. 10.3109/14992029509042213.

Hällgren, M., Larsby, B., and Arlinger, S. (2006). “A Swedish Version of the Hearing in Noise Test (HINT) for Measurement of Speech recognition.” International Journal of

(29)

Hunter, C. R., & Pisoni, D.B. (2018). “Extrinsic Cognitive Load Impairs Spoken Word Recognitionin High- and Low-Predictability Sentences.” Ear and Hearing, 39(2): 378– 389. doi:10.1097/AUD.0000000000000493

Jensen, A. R., Saccuzzo, D. P., & Larson, G. E. (1988). Equating the Standard and Advanced Forms of the Raven Progressive Matrices. Educational and Psychological

Measurement, 48(4), 1091–1095. https://doi.org/10.1177/0013164488484026

Kyllonen P. C., Christal, R. E. (1990). Reasoning ability is (little more than) working-memory capacity?! Intelligence, 14(4), 389–433.

Lunner, T. 2003. “Cognitive Function in Relation to Hearing Aid Use.” International Journal

of Audiology, 42, 49–58. https://doi.org/10.3109/14992020309074624

Michalek, A. M. P., Ash, I. & Schwartz, K. (2018). The independence of working memory capacity and audiovisual cues when listening in noise. Scandinavian Journal of

Psychology, 59, 578–585.

Ng, E. H. N., Rönnberg, J (2020) Hearing aid experience and background noise affect the robust relationship between working memory and speech recognition in noise, International

Journal of Audiology, 59(3), 208-218, DOI: 10.1080/14992027.2019.1677951

Prabhakaran, V., Smith, J. A., Desmond, J. E., Glover, G. H., & Gabrieli, J. D. (1997). Neural substrates of fluid reasoning: an fMRI study of neocortical activation during performance of the Raven’s Progressive Matrices Test. Cognitive psychology, 33(1), 43-63. https://doi.org/10.1006/cogp.1997.0659

Raven, J. (2008). General introduction and overview: The Raven Progressive Matrices Tests: Their theoretical basis and measurement model. In: Raven, J. & Raven, J. (eds.) Uses

and Abuses of Intelligence: Studies Advancing Spearman and Raven’s Quest for Non-Arbitrary Metrics. Unionville, New York: Royal Fireworks Press; Edinburgh, Scotland:

Competency Motivation Project; Budapest, Hungary: EDGE 2000; Cluj Napoca, Romania: Romanian Psychological Testing Services SRL (Chapter 1, p. 17–68).

(30)

Rönnberg, J. (2003) Cognition in the hearing impaired and deaf as a bridge between signal and dialogue: a framework and a model, International Journal of Audiology, 42, 68-76, DOI: 10.3109/14992020309074626

Rönnberg, J., Holmer, E. & Rudner, M. (2019). Cognitive hearing science and ease of language understanding, International Journal of Audiology, 58:5, 247-261, DOI: 10.1080/14992027.2018.1551631

Rönnberg, J., Lunner, T., Ng, E. H. N., Lidestam, B., A. A. Zekveld, P. Sörqvist, … Stenfelt, S. (2016). “Hearing Impairment, cognition and Speech Understanding: Exploratory Factor Analyses of a Comprehensive Test Battery for a Group of Hearing Aid Users, the n200 Study.” International Journal of Audiology 55(11): 623–642. doi:10.1080/14992027.2016.1219775.

Rönnberg, J., Lunner, T., Zekveld, A. A., Sörqvist, P., Danielsson, H., Lyxell, B., … Rudner, M. (2013). The Ease of Language Understanding (ELU) model: theoretical, empirical, and clinical advances. Frontiers in Systems Neuroscience, 7(31).

https://doi.org/10.3389/fnsys.2013.00031

Rönnberg, J., Rudner, M. & Lunner, T. (2011). Cognitive hearing science: The legacy of Stuart Gatehouse. Trends Amplif, 15, 140–148. DOI: 10.1177/1084713811409762.

Stenfelt, S., and Rönnberg, J. (2009). “The Signal-cognition Interface: interactions between Degraded Auditory Signals and Cognitive Processes.” Scandinavian Journal of

Psychology, 50(5): 385–393. doi:10.1111/j.1467-9450.2009.00748.x.

Zekveld, A. A., Kramer, S. E., & Festen, J. M. (2011). Cognitive Load during Speech Perception in Noise: The Influence of Age, hearing Loss, and Cognition on the Pupil Response. Ear and Hearing, 32(4): 498–510. doi:10.1097/AUD.0b013e31820512bb.

Zekveld, A. A., Rudner, M., Johnsrude, I. S., Heslenfeld, D. J., & Rönnberg, J. (2012). Behavioural and fMRI Evidence That Cognitive Ability Modulates the Effect of Semantic Context on Speech Intelligibility. Brain and Language, 122(2): 103–113. doi:10.1016/j.bandl.2012.05.006.

The association between cognition and speech-in-noise perception : Investigating the link between speech-in-noise perception and fluid intelligence in people with and without hearing loss