Skip to main content
NCBI home page
NIHPA Author Manuscripts logoLink to NIHPA Author Manuscripts
. Author manuscript; available in PMC: 2026 Sep 12.
Published in final edited form as: J Acoust Soc Am. 2025 Aug 1;158(2):1367–1376. doi: 10.1121/10.0038978

English band importance functions in competing speech for Spanish-speaking second-language learners of English

Emily Buss 1, Sofia C F Hall 2, Stefani Garcia 2, Grace S Capretta 2, Lauren Calandruccio 2
PMCID: PMC13566126  NIHMSID: NIHMS2204345  PMID: 40844910

Abstract

Band importance functions for speech perception have been well characterized for native speakers of English tested with English speech-in-noise materials. Fewer datasets are available for speech-in-speech recognition or for listeners who are second-language learners of English. The present study evaluated band importance functions for sentences in a two-talker speech background for Spanish/English bilinguals who were second-language (L2) learners of English and two groups of native (L1) English speakers. Target speech was monaural, with three target talkers presented in separate blocks of trials. The L2 speakers and one group of L1 speakers were tested with a diotic masker (MoTm); a second group of L1 speakers was tested with a monaural masker (MmTm). The L2 speakers had higher speech reception thresholds compared to L1 speakers tested with the same stimuli (MoTm), but more similar performance compared to L1 speakers tested in the more challenging condition (MmTm), which lacked binaural difference cues. Compared to both L1 groups, L2 speakers had shallower psychometric functions and benefitted less from semantic context. Band importance functions differed for the three talkers, but they were similar for all three groups of participants, indicating that L1 and L2 speakers largely relied on the same spectral cues for speech-in-speech recognition.

Keywords: speech recognition, bilingual, two-talker speech, spectral weights

I. INTRODUCTION

Natural communication often occurs in multi-source listening environments (Hodgson et al., 2007). Under these listening conditions, two forms of masking are thought to limit access to the target speech: energetic masking, whereby peripheral response to the masker degrades encoding of the target, and informational masking, whereby listeners have trouble differentiating acoustic features of the target and masker (Bronkhorst, 2015). Whereas background speech can exert both types of masking, recognition of speech in a two-talker masker is often limited by informational masking, particularly when speech streams are perceptually similar and there is high stimulus uncertainty (Brungart, 2001). When successful, speech-in-speech recognition at and below 0 dB target-to-masker ratio (TMR) typically relies on “glimpses” of the target that are not energetically masked (Cooke, 2006). Glimpsing requires segregation of the target and masker and selective attention to the spectrotemporal epochs associated with good target cue audibility.

For bilingual listeners, proficiency in the test language affects speech-in-speech recognition (Cooke et al., 2008; Ezzatian et al., 2010; Kilman et al., 2014; Regalado et al., 2019). Under challenging listening conditions, recognition is often better for speech in the listener’s first language (L1) than their second language (L2), particularly if the listener’s L2 was acquired later in life (Regalado et al., 2019). Whereas some datasets indicate that L2 speakers of a target language experience more informational masking than L1 speakers (Cooke et al., 2008), other data indicate similar group effects for energetic and informational masking (Ezzatian et al., 2010). Second language acquisition involves learning new acoustic phonetic cues and developing predictive language models (Garcia Lecumberri et al., 2010; Brodbeck et al., 2024). Poorer masked speech recognition for L2 than L1 speakers of the test language is often described in terms of reduced familiarity with the phonemic, lexical, and syntactic structure of test language, and poorer ability to recognize speech based on spectrotemporally sparse cues compared to the L1 speakers (Calandruccio et al., 2014; Calandruccio and Buss, 2017).

In addition to poorer speech recognition, several studies show greater listening effort for L2 speakers of the test language as compared to L1 speakers. Greater effort when listening in an L2 language has been observed in quiet and in noise (Song and Iverson, 2018; Bsharat-Maalouf et al., 2025), and is seen even when the TMR is adjusted to equate intelligibility for L1 and L2 speakers of the test language (Borghini and Hazan, 2018). Semantic context of the target speech tends to decrease effort and increase performance for individuals listening in their L1, but it can have the opposite effect when speech is in the listener’s L2, increasing effort and decreasing performance (Borghini and Hazan, 2018; Bsharat-Maalouf et al., 2025). Further, proficient L2 speakers’ ability to benefit from semantic context can depend on the quality of the target speech, emerging for clear speech but not for conversational speech (Bradlow and Alexander, 2007). While it is unclear why semantic cues might negatively impact sentence recognition, inefficient use of context in L2 speakers has been attributed to reduced cognitive resources when listening to an unfamiliar language (see also Golestani et al., 2009).

Given the fundamental differences between masked speech perception for L1 and L2 speakers of the test language, it is possible that L2 speakers could weight target speech cues distributed across frequency differently than L1 speakers. Native speakers of English tend to rely on speech cues in the spectral region of 2 kHz more than cues below or above that frequency, although functions differ in detail depending on the stimulus conditions, including the choice of corpora and target talkers (ANSI, 1997; Yoho et al., 2018) and the presence of background noise (Apoux and Bacon, 2004). Band Importance Functions (BIFs) also differ in noise and competing speech (Buss and Bosen, 2021; Bosen et al., 2024), and for masker talkers who are facing towards vs. away from the listener (Ananthanarayana et al., 2024). Most estimates of band importance are based on data from native English speakers listening to English target speech. There are also data indicating that BIFs can differ significantly as a function of the test language (Wong et al., 2007; Jin et al., 2016; Du et al., 2019), but we are not aware of any data on possible effects of language experience on BIFs among L1 and L2 speakers of English.

There are numerous reasons why L2 speakers might weight spectral information differently than L1 speakers for speech-in-speech recognition. A few examples are described here. First, early auditory experience has an outsized influence on perception of phonemic contrasts, with exposure to language early in infancy increasing sensitivity to contrasts that are meaningful in that language (Kuhl et al., 2008). Although phonetic boundaries continue to be shaped by language input beyond that period (Kutlu et al., 2024), acquiring a language later in life could result in less precise or less robust representation of the phonemes in that language, such that L2 speakers require more information or different information than L1 speakers to discriminate those phonemes (Iverson et al., 2003). Second, limited proficiency in the test language is thought to delay lexical access, increasing demands on memory and executive function. Those resources are also required for segregating simultaneous streams of speech and selectively attending to the target (Koelewijn et al., 2012; Tamati et al., 2013). These two processes – lexical access and segregation – might over-tax cognitive resources for L2 speakers and increase the number or quality of cues required for recognition. Third, differences in the way that L1 and L2 speakers use semantic context could affect band importance. The amount of semantic context available in the target speech affects band importance among monolingual speakers of English (e.g., Studebaker and Sherbecoe, 1991; DePaolis et al., 1996) and Mandarin (Du et al., 2019), with broader functions and greater low frequency weight for high-context than low-context speech. Second-language learners are less adept than native speakers at using semantic context to support recognition (Bradlow and Alexander, 2007; Shi and Koenig, 2016), and their BIFs may reflect that. Whereas failure to make use of semantic context might result in an increased reliance on cues in the region of 2 kHz, underdeveloped phonemic templates and slowed lexical access may be associated with a need for more information distributed across frequency (Calandruccio et al., 2014; Calandruccio and Buss, 2017). These alternatives are not independent or mutually exclusive, and this is not an exhaustive list.

The present study estimated BIFs for L1 and L2 speakers of English who were performing an English speech-in-speech recognition task. BIFs were computed based on word-level sentence recognition data when the overall TMR was set to the listener’s speech reception threshold (SRT). Two groups of participants were tested with masker speech presented diotically to both ears (Mo) and target speech presented monaurally (Tm); one group included L1 speakers of English, and the other included L2 speakers of English. The MoTm stimulus creates a binaural difference cue that helps listeners segregate the target and masker speech. The binaural cue was expected to benefit both groups of participants based on prior data showing that perceived spatial separation supports stream segregation for L1 and L2 speakers of the test language, with comparable magnitudes of binaural benefit (Ezzatian et al., 2010). Based on prior research, performance was expected to be poorer for L2 speakers than L1 speakers in the MoTm stimulus condition. This expectation prompted the inclusion of a third participant group of L1 speakers who were tested in a more challenging stimulus condition, with both the masker and target speech presented monaurally to one ear (MmTm). Performance for the L1 English speakers in this more challenging condition was expected to approximately match L2 English speakers’ performance in the easier (MoTm) condition. Comparing speech recognition for L1 and L2 speakers of English in conditions that are matched either with respect to the stimulus or with respect to performance allowed us to evaluate effects related to differences in overall TMR or differences in cue audibility.

II. METHODS

A. Participants

This study recruited young adults with normal hearing, defined as pure-tone thresholds ≤ 20 dB HL at octave frequencies 250–8000 Hz bilaterally (ASHA, 2005). There were three groups, defined by the participants’ language histories and the stimulus conditions used to test them. The first group included L2 English speakers, all Spanish/English bilinguals from Spain who were in the United States as part of a study abroad program. This group included 15 individuals, 18–22 years of age (13 females, 2 males). Race was reported as White (n=13), Asian (n=1)1, or mixed race (n=1), and all but one participant identified as Hispanic. This group reported starting to learn English between 3 and 8 years of age (M = 4.7 yrs) and residing in the United States between 1 and 15 months (M = 5.4 mos), with all but two living in the US for less than one year. Language proficiency for L2 speakers of English was assessed with the Bilingual Language Profile (BLP; Birdsong et al., 2012). In addition to Spanish and English, 12 participants in this group spoke one or more additional languages, including Galician, Italian, French, German, and Japanese. After data collection, it was determined that participants consistently misunderstood one of the questions on this instrument; this question was removed and weights applied to each response were adjusted to maintain the original balance of questions across domains.2 Resulting scores on the BLP were between −149 and −57 (M = −90.4). As expected, these data indicated that all participants in this group were dominant speakers of Spanish. For reference, scores on this instrument can range between −218 (maximal dominance in Spanish) and 218 (maximal dominance in English), whereas a score of 0 would indicate a truly balanced bilingual.

There were two groups of L1 speakers of American English, each with 10 participants. The first group was tested in the same stimulus conditions as the L2 English speakers (MoTm) and is referred to below as the L1 stimulus-matched group. Participants in this group were 19–23 years of age (8 female, 1 male, 1 non-binary). Race was reported as White (n=8), Black (n=1), or Asian (n=1), and only one participant identified as Hispanic. All were monolingual speakers of English. The second group of L1 speakers of English was tested in a more challenging condition than the other two groups (MmTm). Race was reported as White (n=8), Black (n=1), or more than one race (n=1), and only one participant identified as Hispanic. In this group, 8 of 10 participants were monolingual English speakers; one participant was a late learner of Spanish, and one was a late learner of Mandarin3. The rationale for using a more challenging condition for this group of L1 speakers of English was to approximately match performance achieved by the L2 English speakers. This group is therefore referred to as the L1 performance-matched group.

B. Stimuli

The target speech for this experiment was from the Basic English Lexicon (BEL; Calandruccio and Smiljanić, 2012) corpus of sentences. This corpus uses words spontaneously produced by second-language learners of English in interviews on conversational topics. These sentences were recorded by three female talkers. Target talker JC was recorded at 44.1 kHz and had a mean fundamental frequency (F0) of 179 Hz. Target talkers F1og and F2og were recorded at 22.05 kHz and had mean F0s of 187 and 193 Hz, respectively. All three of these recordings were originally made for previous projects (Calandruccio and Smiljanić, 2012; O'Neill et al., 2020). The present study used four lists of 25 BEL sentences recorded by each talker; those lists were 1–4 for talker JC, 5–8 for talker F1og, and 9–12 for talker F2og. The rationale for testing speech produced by multiple target talkers was the observation of large differences in speech-in-speech recognition for different target-masker combinations (Freyman et al., 2007; Iyer et al., 2010; Calandruccio et al., 2017).

Masker speech was produced by two different female talkers reading short narratives, with mean F0s of 183 and 205 Hz. Recordings were made at 44.1 kHz, and pauses > 200 ms were manually removed. These stimuli were used previously in a masked speech perception experiment reported by (Wasiuk et al., 2022). The two recordings were 192 and 194 sec in duration. The two streams were equated in root-mean-square (RMS) level and summed.

The target and masker stimuli that were recorded at 44.1 kHz were downsampled to 22.05 kHz to match the stimuli recorded at that lower sampling rate. Each target sentence was temporally centered in a randomly selected segment of the two-talker masker that was 1-sec longer than the target. These stimuli were saved to disk as WAV files. The long-term average speech spectra for the targets and masker are shown in Figure 1.

Figure 1.

Figure 1.

Open in a new tab

The long-term average speech spectra (LTASS) for each target talker and the two-talker masker. Line style is defined in the legend.

C. Procedures

For L2 and L1 stimulus-matched groups, the masker was presented diotically in both ears, and the target was presented monoaurally to the right ear (MoTm), providing a binaural difference cue. For the L1 performance-matched group, the masker and the target were presented monaurally to the right ear (MmTm). All participants completed 3 blocks of trials, with 100 sentences per block. Each block comprised target sentences from one of the three target talkers. The order of sentences within a block and the order in which participants completed blocks with different target talkers were randomized. The first 50 sentences of each block were presented in two interleaved adaptive tracks. Both tracks used a 1-down, 1-up stepping rule, but with different criteria. One track used a lax criterion, where the TMR was reduced following trials with one or more words correct. The other track used a strict criterion, where the TMR was reduced following trials with all or all but one word correct. This procedure ensures that the first 50 trials include TMRs below and above the TMR associated with 50% correct keyword recognition. Data from these 50 trials were fitted with a logit to characterize the psychometric function (EQ1.)

y=11+exp-x−αβ (EQ1.)

where y is the proportion correct, x is the target level in TMR, α is the SRT associated with 50% correct, and β is the function slope. The remaining 50 sentences in the block were presented at the SRT estimated based on the function fit. Data from the final 50 trials were used to estimate BIFs for individual participants and conditions.

The method for estimating BIFs entailed passing each stimulus through a gammatone filterbank, with one filter per equivalent rectangular bandwidth (ERB; Moore and Glasberg, 1983) between the 4th and 36th ERB (center frequencies of 123 and 9,657 Hz). The time points associated with the beginning and end of each keyword were identified with Praat (Boersma and Weenink, 2024) and used to compute the mean TMR by keyword. Weights were then estimated as the correlation between TMR and correct keyword recognition for each frequency band. More information about this method can be found in recent papers from our group (Buss and Bosen, 2021; Bosen et al., 2024). The mean and the 95% confidence interval for group BIFs in each condition were computed with the Fisher’s z-transformation of correlation (weight) to normalize error variance.

D. Statistical analysis

Analyses were carried out with R (R Core Team, 2025) using the lme4 (Bates et al., 2015) and emmeans (Lenth, 2025) libraries. For models evaluating the SRTs, psychometric function slopes, and recognition accuracy, the three participant groups and three talkers were treated as categorical variables, and there was a random intercept for each participant. F-tests of marginal means were used in the text to describe results of linear mixed-effects models. Full model results are available in supplemental tables (Buss, 2025). For these models, the reference condition was L2 speakers of English tested with talker JC. For the models evaluating BIFs, the reference frequency was the 22nd ERB (center frequency 2,212 Hz), which is close to the band associated with peak importance in previous data.

III. RESULTS

A. Psychometric function fits to adaptive data

Data from the first 50 trials of each track resulted in generally well-fitted psychometric functions, with a median r2 value of 0.92 (0.61 to 0.99 across participants and talkers). To account for the possibility that some participants, particularly L2 speakers, may not achieve 100% accuracy even at favorable TMRs, we introduced a third psychometric function parameter that allowed the function to asymptote below 100%. Introducing this third psychometric function parameter increased the median percent of variance accounted for by less than one percentage point on average. The increase in percent of variance accounted for by a 3-parameter fit exceeded five percentage points in only three datasets, all for L2 speakers of English and talker JC. This result suggests that those three participants in the L2 group might have scored below 100% correct for talker JC even in quiet. The results reported below are based on 2-parameter function fits, but repeating the analyses using 3-paremeter fits does not change the general pattern of results. Panels A1-A3 of Figure 2 show psychometric functions based on median parameter values for each talker and participant group. Distributions of SRT and psychometric function slope for individual participants are shown in Panels B and C, respectively. Fits for individual participants are shown in the repository for this project (Supplemental Figure 1; Buss, 2025).

Figure 2:

Figure 2:

Open in a new tab

Results of adaptive testing. Median psychometric function fits are shown separately for each talker and participant group (panels A1-A3). Line color reflects participant group, as defined in the legend. Panel B shows distributions of SRT (α) estimated based on function fits to individual participants’ data. Results are shown separately by talker, indicated on the abscissa, and participant group, indicated by color. Panel C shows distributions of psychometric function slopes (β) fitted to indivdiual participants’ data. Horizontal lines indicate the medians, boxes span 25th-75th percentiles, and vertical lines indicate the range of values apart from outliers, which are indicated with circles.

As expected, the SRT for L2 speakers of English was higher than for L1 speakers when both groups were tested in the same stimulus condition (MoTm). For participants tested in this condition, the mean SRT was 2.5 dB higher for L2 speakers than L1 speakers (2.2 – 2.9 dB across talkers). Performance was nearly matched when comparing L2 speakers tested with the binaural cue (MoTm) and L1 speakers tested without that cue (MmTm); for these conditions, the mean difference in values of SRT range from −1.6 dB (talker JC) to 0.7 dB (talker F2og). A linear mixed effect model of SRT indicates significant effects of participant group (F2,32 = 10.32, p < 0.001) and talker (F2,64 = 61.54, p < 0.001), and a non-significant trend for an interaction between group and talker (F4,64 = 2.39, p = 0.060). Model coefficients are reported in the repository for this project (Table I; Buss, 2025). For data from L2 speakers, there was not a significant correlation between SRT and the BLP score, with correlations of r = −0.35 to 0.07 (p = 0.196 to 0.927).

The mean psychometric function slope was shallower (values of β were larger) for L2 speakers of English (β = 3.1) than for L1 speakers tested in the MoTm (β = 2.4) and MmTm (β = 2.6) conditions. This observation was confirmed with a linear mixed effect model indicating a significant effect of group (F2,32 = 4.56, p = 0.018). The effect of talker failed to reach significance (F2,64 = 1.06, p = 0.354), as did the two-way interaction (F4,64 = 0.38, p = 0.820). Model coefficients are reported in the repository for this project (Table II; Buss, 2025). One practical consequence of group differences in function slopes is that the group differences in SRT depend on the criterion level used to define that threshold. For example, for data obtained in the MoTm condition with talker JC, the difference in performance for L2 and L1 performance-matched speakers of English based on group median functions was 2.9 dB when SRTs were defined at 50% and 4.3 dB at 90%.

B. Fixed-TMR data

Data from the second 50 trials of each track indicated similar percent correct word recognition across talkers and participant groups, with mean scores of 49–57% correct. A linear mixed effect model assessing logit-transformed proportion correct scores indicated non-significant effect of participant group (F2,32 = 0.98, p = 0.388), but scores with talker JC were lower than for the other two talkers (F2,68 = 3.30, p = 0.043). Model coefficients are reported in the repository for this project (Table III; Buss, 2025). This shows that individualized TMRs based on the SRT obtained in the first 50 trials were generally successful at normalizing word recognition performance between participant groups in the fixed-TMR trials, although it is unclear why scores would be lower for talker JC than the other two talkers. In addition to word-level scores, data from the fixed-TMR trials were also evaluated using sentence scoring, where all keywords must be reported correctly to qualify as a correct response. For sentence scoring, L2 speakers achieved a mean score of 20% correct, whereas the two groups of L1 speakers both had mean scores of 25% correct. Boothroyd and Nittrouer (1988) proposed using the relationship between the probability of correct word recognition (Pword) and sentence recognition (Psent) to characterize the benefit provided by sentence context using the j factor, where Psent = Pwordj. Theoretically, values of j can range between 1 and the total number of keywords in each sentence (in this case 4). Smaller values of j indicate greater use of context. In the present dataset, mean values of j by group and talker ranged from 1.9 to 2.7. A linear mixed effects model, with the context parameter j as the dependent variable, indicates a significant effect of participant group (F2,32 = 4.17, p = 0.025), reflecting poorer use of context (higher values of j) for L2 speakers than L1 speakers, and a significant effect of talker (F2,68 = 55.20, p < 0.001), reflecting lower values of j for talker JC than the other two talkers. Model coefficients are reported in the repository for this project (Table IV; Buss, 2025). Values of j tend to increase with increasing percent correct word recognition (Boothroyd and Nittrouer, 1988; Smits and Zekveld, 2021), and the observation of lower values of j for talker JC than the other two talkers could be related to the poorer word recognition scores for this talker. An additional consideration when evaluating these results is the fact that access to the target speech cues necessary for recognition can vary markedly across words for speech-in-speech recognition, and this variability can affect the resulting values of j (Smits and Zekveld, 2021). More work is needed to fully understand these effects, but it is likely that they impact results of L1 and L2 participants similarly.

Figure 3 shows spectral weights plotted as a function of frequency (BIFs) for each talker and participant group. There is marked variability in results between the 10–15 participants in each group and condition; the median standard error of weights across frequency is between r = 0.015 and 0.022. The variability observed between individual participants is approximately an order of magnitude less than the range of weights observed across frequency, which is approximately 0.3 (between r ≈ −0.1 and 0.2). Despite variability between participants, visual inspection of the data suggests that there are differences in weights between talkers. For example, mean weights for talker JC are relatively flat between 500 and 4000 Hz, but there is a local minimum around 1000 Hz for talker F1og. Of particular interest in the present study, functions are broadly similar for the three participant groups. Whereas mean weights for L2 speakers fall below those for L1 speakers for talker F2og between 750 and 4000 Hz, that difference is modest (max of 0.055) relative to variability across participants – less than two times the standard error of the mean even without accounting for multiple comparisons.

Figure 3:

Figure 3:

Open in a new tab

Weights estimated by frequency for each talker based on fixed-TMR data. Fits to individual participants’ data are indicated with grey lines, mean fits by talker and group are shown in thick colored lines, and shading shows the associated 95% confidence intervals. Results are shown separately by talker, as indicated with column titles, and by group, as indicated in the ordinate labels. The mean TMRs across participants for each talker and group are indicated in the lower right of the associated panel. The bottom row of panels shows mean functions for all three groups plotted together.

These visual impressions of BIFs were evaluated with a linear mixed effects model predicting weight as a function of frequency, talker, and group. For this model, weight was represented as a z-score. Results evaluated with a marginal F-test appear in Table 1, and model coefficients are reported in the repository for this project (Table V; Buss, 2025). There were significant main effects of frequency (p < 0.001) and talker (p < 0.001), and a non-significant effect of group (p = 0.085). There were significant interactions between frequency and talker (p < 0.001) and between frequency and group (p = 0.006), but not between talker and group (p = 0.802). The three-way interaction also missed significance (p = 0.062).

Table 1:

Marginal ANOVA evaluating statistical model of weights that included three categorical variables and their interactions: frequency (33 levels), talker (3 levels), and group (3 levels).

numDF denDF F-value p-value
(Intercept) 1 3136 110.28 <0.001
Freq 32 3136 16.50 <0.001
Talker 2 3136 8.69 <0.001
Group 2 32 2.66 0.085
Freq:Talker 64 3136 12.60 <0.001
Freq:Group 64 3136 1.51 0.006
Talker:Group 4 3136 0.41 0.802
Freq:Talker:Group 128 3136 1.20 0.062
Open in a new tab

When the model described above is run with just the two sets of L1 data (MoTm and MmTm), then the main effect of group and all interactions with group fail to reach significance (p ≥ 0.339). This implicates differences between L1 and L2 speakers of English in effects of group that were observed in the full model, as opposed to differences in the stimulus conditions used to test them. Based on this result, the interaction between frequency and group was evaluated with a model that coded group solely based on language experience (L1 or L2). For this model, the estimated marginal means indicate significant differences (p < 0.05) between weights for L1 and L2 speakers of English for 5–10 bands for each talker. The bands associated with group differences differ between talkers, however; for a criterion of p < 0.05, there is no single frequency band for which a significant effect is observed for all three talkers. For talkers F1og and F2og, all significant differences are associated with lower weight for L1 than L2 speakers. For talker JC, significant differences are associated with both higher and lower weight for L1 speakers. For more details, see the repository for this project (Table VI; Buss, 2025). The effect size associated with differences in BIFs for L1 and L2 speakers was quantified by comparing models with and without that factor and associated interactions; the conditional R2 values fall from 0.61 to 0.59 with omission of group from the model. This result is consistent with the conclusion that about 2% of variance in the BIFs reported here can be attributed to the participant’s language background.

Contrasts were also used to evaluate the interaction between talker and frequency in the full model. Pairwise comparisons among the three talkers indicate significant differences (p < 0.05) between talkers for 20–24 of the 33 bands for L2 participants and 19–24 of the 33 bands for L1 participants (data for MoTm and MmTm combined). The only frequency bands for which no significant differences are observed for any pair of talkers in either language group are ERBs 20 and 21 (1636–2084 Hz). For more details, see the repository for this project (Table VII; Buss, 2025). The effect size associated with differences in BIFs between talkers was quantified by comparing models with and without that factor and associated interactions; the conditional R2 values fall from 0.61 to 0.44 with omission of talker from the model. This result is consistent with the conclusion that about 17% of variance in the BIFs reported here can be attributed to the talker.

IV. DISCUSSION

This study was designed to compare BIFs for speech-in-speech recognition in L1 and L2 speakers of English. All participants were young adults with normal hearing, and the L2 speakers were Spanish-dominant Spanish/English bilinguals. The L2 speakers were tested with a diotic masker and a monaural target (MoTm); this condition provides a binaural difference cue that was intended to improve performance relative to a fully monaural or fully diotic presentation, as binaural difference cues have been shown to improve performance for both monolingual and bilingual listeners (Ezzatian et al., 2010). There were two reasons for testing L2 participants with a binaural difference cue. First, informational masking tends to be greater at negative than positive TMRs (Freyman et al., 2008; Swaminathan et al., 2015), and we wanted to test performance based on glimpsing of speech. This objective was met. Mean SRTs in the MoTm condition for L2 speakers were negative for all three talkers, with values ranging from −4.9 to −9.7 dB TMR. Second, we wanted to run a group of L1 participants in a more challenging condition, for a performance-matched comparison with the L2 participant data. That more challenging condition used a monaural masker and monaural target (MmTm). This condition was successful at approximately matching performance across listener groups, as described below. For each participant and target talker, the first phase of testing identified an SRT, and the second phase of testing was carried out at that TMR. Results from each phase of testing are discussed in turn.

A. Psychometric function fits to adaptive data

For the MoTm condition, the mean SRT was 2.5 dB higher for L2 speakers than L1 speakers of English. This result can be compared to data from Regalado et al. (2019). That study evaluated English sentence recognition in a two-talker masker, with all female talkers and stimuli presented diotically. Listeners were young adults who were monolingual or Mandarin-dominant Mandarin/English bilingual speakers of English.4 Five of the 20 L2 listeners from the Regalado et al. study had similar lengths of US residency as the L2 participants in the present study (< 1 year), but on average participants from the prior dataset had resided in the US for 2.7 years (0–7 years). The L2 groups in both studies reported beginning to learn English between 3 and 10 years of age. Mean SRTs for speech-in-speech recognition reported by Regalado et al. (2019) were −4.8 and −2.3 dB TMR for L1 and L2 groups, respectively. These SRTs can be compared to the mean SRTs across talkers for the MoTm condition in the present study, which were −10.1 and −7.5 TMR for L1 and L2 groups, respectively. The lower mean SRTs compared to the data of Regalado et al. (2019) are not surprising given the binaural difference cue available for these conditions, but the group effect was similar across studies, with values near 2.5 dB (see also: Ezzatian et al., 2010).

Testing L1 speakers in the MmTm condition, without the binaural difference cue, was intended to approximately match performance for the L2 group. Compared to L1 speakers tested in the MoTm condition, removing the binaural difference cue increased mean SRTs by 3.8 dB (JC), 3.0 dB (F1og) and 2.2 dB (F2og). This can be compared to a mean binaural benefit of 6.8 dB observed for sentences in two-talker speech, with both the target and masker diotic (MoTo) or with the target diotic and the masker out of phase at the two ears (e.g., MoTpi; Calandruccio et al., 2020). The goal of matching L1 and L2 performance by removing the binaural difference cue was somewhat successful. For talkers F1og and F2og, mean SRTs were within approximately 0.5 dB for L2 speakers tested with the binaural difference cue and L1 speakers tested without the binaural difference cue. That performance match was poorer for talker JC, with the mean SRT for L2 speakers falling 1.6 dB below the mean for L1 speakers of English.

While SRT was the primary parameter of interest from the psychometric function fits, the function slope was also evaluated. The psychometric function slope characterizes the rate at which recognition accuracy increases with increasing TMR. Function slope was shallower for L2 speakers than for either group of L1 speakers. This finding is consistent with some prior data. For example, Mayo et al. (1997) tested sentence recognition in a babble masker and observed shallower psychometric function slopes for L2 as compared to L1 speakers of the test language. However, differences in function slope are not always observed. Ezzatian et al. (2010) reported steeper psychometric function slopes for L1 than L2 speakers tested in speech-shaped noise, but similar slopes when the masker was two-talker speech. Shallower slope for L2 speakers could indicate a wider pattern of lexical activation and slower elimination of competitors, due to less precise representation of the statistical properties of the target language.

B. Fixed-TMR data

As expected, performance hovered around 50% correct when participants were tested at the TMR corresponding to their SRT. The pattern of responses (keywords correct vs. incorrect) was used to estimate BIFs for each participant and each target talker. These functions differed for the three target talkers, with talker accounting for approximately 17% of variance in the data. This was not expected given previous data on speech-in-noise, which provides little evidence of reliable differences between talkers apart from differences related to sex (Yoho et al., 2018). It is possible that talker effects are more pronounced for speech-in-speech than speech-in-noise, or that mismatches in the long-term speech spectra of the target and masker speech affect weights.

Despite variability across talkers, mean BIFs were generally similar across groups of participants. Whereas L2 speakers applied significantly less weight to cues in some frequency bands compared to L1 speakers, this effect was small, accounting for only 2% of variance in the data. BIFs for two groups of L1 participants – the stimulus-matched group (tested with a binaural difference cue; MoTm) and the performance-matched group (tested without a binaural difference cue; MmTm) – did not differ. A priori, it was not clear that BIFs would be functionally the same for L1 speakers tested in the two conditions. Bosen et al. (2024) observed a difference in BIFs depending on the spatial distribution of target and masker speech. When the sources were separated on the horizontal plane, listeners applied relatively more weight to low frequencies compared to when sources were co-located. Bosen et al. (2024) speculated that these differences could be related to poorer high-frequency audibility at TMRs associated with ~50% correct when sources are separated in space5. While the present study used headphone presentation, the binaural difference cue in the MoTm condition produces differences in perceived lateralization, whereas the MmTm condition would create the perceptual impression of the target and masker being co-located to one side. We might therefore have expected to see more low-frequency weight for L1 participants tested in the MoTm condition than the MmTm. The fact that we did not observe a difference in BIFs could be related to consistency of stimulus features in the (right) target headphone in the two conditions (as compared to loudspeaker presentation used in Bosen et al., 2024), or to the relatively high target level even in the easier (MoTm) condition.

Data collected at a fixed TMR were also used to evaluate participants’ use of sentence context. This analysis compared the distribution of percent correct scores by word and by sentence. High reliance on context would result in more instances of all words in a sentence being correct or incorrect than predicted based on the overall probability of correct word recognition. Results indicated that L1 speakers were more efficient at utilizing semantic context than L2 speakers. This finding is consistent with some prior research (Mayo et al., 1997; Golestani et al., 2009). For example, Mayo et al. (1997) observed a larger group difference between L1 and L2 speakers for high-context than low-context sentences. Interestingly, Bradlow and Alexander (2007) observed that L2 speakers of the test language did not benefit from sentence context when speech was produced in a conversational speaking style, but they could benefit from context when talkers used a clear speaking style. This suggests that the cognitive resources necessary for utilizing context may be particularly taxed when L2 speakers are faced with unfamiliar or imprecise phonemic information. Among L1 speakers, psychometric functions are shallower for sentences with low levels of semantic context as compared to high context. This is explained in terms of the lower requirement for acoustic cue audibility when there are context cues available. It is therefore possible that group differences in psychometric function slope observed in the present dataset are related to the ability to make efficient use of sentence context. However, this association was not observed within data for L2 speakers. Psychometric function slopes estimated for each talker based on adaptive data were not significantly correlated with the values of j estimated based on fixed-TMR data (r = −.20 to .08, p ≥ .482). The homogeneity of the L2 speakers’ English language experience may have undermined our ability to observe an association between these factors in the present dataset.

C. Other considerations

The present study included L1 speakers of American English and Spanish-dominant Spanish/English bilinguals from Spain who were L2 speakers of English and were attending a study abroad program in the US at the time of the experiment. The BIFs measured for these two populations were broadly similar, but this does not rule out larger differences in other populations. For example, less proficient L2 speakers of English who have less current and daily exposure to English might be more likely to weight spectral cues differently. It is also possible that L1 speakers of other languages might apply different BIFs when listening to English. English and Spanish are both Indo-European languages rooted in Latin, meaning that they share many phonemic qualities, such as open, unobstructed vowels, as well as common prefixes, suffixes, and root words (D’elattre, 1964; Baldi and Cuzzolin, 2018). In contrast, there are marked phonetic, lexical, and prosodic differences between English and languages like Korean or Chinese (e.g., Tang et al., 2025), the languages for which prior data indicate differences in BIFs compared to English (Wong et al., 2007; Jin et al., 2016; Du et al., 2019). Further exploration of BIFs for L2 speakers of other languages, particularly those that differ markedly from English, is warranted.

Another feature of the present experiment that could bear on the results is the focus on speech recognition performance at threshold. When asked to adjust background noise to a level that is acceptable when listening to a long story, listeners with normal hearing who are L1 speakers of the test language tend to indicate positive TMRs (e.g., 10 dB TMR; reivewed by Eddins et al., 2013), and when asked to indicate the highest background level they are willing to tolerate for extended periods of listening, they tend to select levels that correspond to near perfect recognition (Recker and Micheyl, 2017). Data obtained at more challenging TMRs and lower levels of performance may not fully capture the differences that L1 and L2 speakers face in natural listening conditions, such as the listening effort required to support that performance or reductions in comprehension or memory that can be associated with difficult listening conditions (Baese-Berk et al., 2023). While existing methods for estimating BIFs do not capture these aspects of speech perception, it is possible that the spectral distribution of speech cues could impact higher-level features of auditory communication.

V. CONCLUSIONS

The present study evaluated English speech-in-speech recognition for two groups of native-English speaking listeners and a group of Spanish-dominant Spanish/English bilinguals who grew up in Spain, began learning English between 3 and 8 years of age (M = 4.7 yrs), and had minimal experience immersed in an English-speaking culture. Consistent with prior literature reviewed in the introduction, this group of L2 speakers had higher SRTs and shallower psychometric function slopes, and took less advantage of semantic context than their L1 counterparts. Differences in BIFs for L1 and L2 speakers were observed, but those effects accounted for only 2% of variance. Perhaps the more striking result is differences in BIFs between talkers, which accounted for 17% of variance. More research is needed to understand the factors affecting BIFs for different talkers and listeners, including L1 speakers of languages from different language families.

Supplementary Material

Data Availability

Data, supplemental tables, and the supplemental figure can be accessed at https://doi.org/10.17605/OSF.IO/T6ZV5.

Acknowledgments

Financial support for this project was provided by the College of Arts and Sciences at Case Western Reserve University. Adam Bosen provided helpful feedback on this project. RedCAP support for our linguistic questionnaires was provided by Clinical and Translational Science Award number UM1 TR004528.

Footnotes

1.

The Spanish/English bilingual participant who identified as Asian was adopted from China and raised in Spain.

2.

Based on comparison to demographic and language experience information collected from participants in the L2 English group, as well as discussion after data collection, there was evidence that a number of participants misunderstand question #3 on the BLP. That question asks, “How many years of classes (grammar, history, math, etc.) have you had in the following languages (primary school through university)?” and provides response alternatives (0–20+) for English and Spanish. Participants’ responses suggest that they were interpreting this question as asking about years of instruction in English as a foreign language. Omitting responses to this question changed the overall dominance score by −6.8 to 1.2 points (M = −3.6, SD = 2.0).

3.

The L1 participant who was Spanish/English bilingual started learning Spanish at 14 years of age, reported being comfortable conversing in Spanish, but used Spanish in a very limited range of settings. The L1 participant who was Mandarin/English bilingual started hearing Mandarin at 15 years of age, was comfortable conversing in Mandarin, but rarely used Mandarin outside of classroom instruction.

4.

Regalado et al. (2019) also tested simultaneous English dominant Mandarin/English bilinguals. Results for this group were very similar to results for monolingual English speakers.

5.

It is possible that reduced high-frequency audibility may have played a role in the broader and flatter BIF for talker JC than for talkers F1og and F2og. As illustrated in Figure 1, the long-term average speech spectrum for talker JC contained less energy above 4 kHz than for those other two talkers, with a maximum difference of 8–9 dB near 7 kHz.

Author Declarations

None of the authors have a conflict of interest to disclose. Human participants provided written informed consent, and the research protocol was approved by the IRB of Case Western Reserve University (IRB-2015–1195).

References

  1. Ananthanarayana RM, Buss E, and Monson BB (2024). “Band importance for speech-in-speech recognition in the presence of extended high-frequency cues,” J. Acoust. Soc. Am. 156, 1202–1213. [DOI] [PMC free article] [PubMed] [Google Scholar]
  2. ANSI (1997). ANSI S3.5–1997 (R2017), American National Standard Methods for Calculation of Speech Intelligibility Index (American National Standards Institute, New York: ). [Google Scholar]
  3. Apoux F, and Bacon SP (2004). “Relative importance of temporal information in various frequency regions for consonant identification in quiet and in noise,” J. Acoust. Soc. Am. 116, 1671–1680. [DOI] [PubMed] [Google Scholar]
  4. ASHA (2005). “Guidelines for Manual Pure-Tone Threshold Audiometry,” in ASHA Practice Policy (American Speech-Language-Hearing Association, Rockville, MD: ). [Google Scholar]
  5. Baese-Berk MM, Levi SV, and Van Engen KJ (2023). “Intelligibility as a measure of speech perception: Current approaches, challenges, and recommendations,” J. Acoust. Soc. Am. 153, 68–76. [DOI] [PubMed] [Google Scholar]
  6. Baldi P, and Cuzzolin P (2018). “Indo-European Languages,” in The World’s Major Languages, edited by Comrie B (Routledge, London: ), pp. 23–50. [Google Scholar]
  7. Bates D, Martin M, Bolker B, and Walker S (2015). “Fitting Linear Mixed-Effects Models Using {lme4},” J. Stat. Softw. 67, 1–48. [Google Scholar]
  8. Birdsong D, Gertken LM, and Amengual M (2012). “Bilingual language profile: An easy-to-use instrument to assess bilingualism.,” (https://sites.la.utexas.edu/bilingual/, COERLL, University of Texas at Austin; ). [Google Scholar]
  9. Boersma P, and Weenink D (2024). “Praat: Doing phonetics by computer,” (www.praat.org). [Google Scholar]
  10. Boothroyd A, and Nittrouer S (1988). “Mathematical treatment of context effects in phoneme and word recognition,” J. Acoust. Soc. Am. 84, 101–114. [DOI] [PubMed] [Google Scholar]
  11. Borghini G, and Hazan V (2018). “Listening Effort During Sentence Processing Is Increased for Non-native Listeners: A Pupillometry Study,” Front. Neurosci. 12, 152. [DOI] [PMC free article] [PubMed] [Google Scholar]
  12. Bosen AK, Wasiuk PA, Calandruccio L, and Buss E (2024). “Frequency importance for sentence recognition in co-located noise, co-located speech, and spatially separated speech,” J. Acoust. Soc. Am. 156, 3275–3284. [DOI] [PubMed] [Google Scholar]
  13. Bradlow AR, and Alexander JA (2007). “Semantic and phonetic enhancements for speech-in-noise recognition by native and non-native listeners,” J. Acoust. Soc. Am. 121, 2339–2349. [DOI] [PubMed] [Google Scholar]
  14. Brodbeck C, Kandylaki KD, and Scharenborg O (2024). “Neural Representations of Non-native Speech Reflect Proficiency and Interference from Native Language Knowledge,” J. Neurosci. 44, e0666232023. [DOI] [PMC free article] [PubMed] [Google Scholar]
  15. Bronkhorst AW (2015). “The cocktail-party problem revisited: Early processing and selection of multi-talker speech,” Atten. Percept. Psychophys. 77, 1465–1487. [DOI] [PMC free article] [PubMed] [Google Scholar]
  16. Brungart DS (2001). “Informational and energetic masking effects in the perception of two simultaneous talkers,” J. Acoust. Soc. Am. 109, 1101–1109. [DOI] [PubMed] [Google Scholar]
  17. Bsharat-Maalouf D, Schmidtke J, Degani T, and Karawani H (2025). “Through the Pupils' Lens: Multilingual Effort in First and Second Language Listening,” Ear Hear. 46, 494–511. [DOI] [PubMed] [Google Scholar]
  18. Buss E (2025). “English band importance functions in competing speech for spanish-speaking second-language learners of English: Data,,” Open Science Framework, 10.17605/OSF.IO/T17606ZV17605. [DOI] [PMC free article] [PubMed] [Google Scholar]
  19. Buss E, and Bosen AK (2021). “Band importance for speech-in-speech recognition,” JASA Express Lett. 1, 084402. [DOI] [PMC free article] [PubMed] [Google Scholar]
  20. Calandruccio L, and Buss E (2017). “Spectral integration of English speech for non-native English speakers,” J. Acoust. Soc. Am. 142, 1646–1654. [DOI] [PMC free article] [PubMed] [Google Scholar]
  21. Calandruccio L, Buss E, and Bowdrie K (2017). “Effectiveness of two-talker maskers that differ in talker congruity and perceptual similarity to the target speech,” Trends Hear. 21, 2331216517709385. [DOI] [PMC free article] [PubMed] [Google Scholar]
  22. Calandruccio L, Buss E, and Hall JW (2014). “Effects of linguistic experience on the ability to benefit from temporal and spectral modulation,” J. Acoust. Soc. Am. 135, 1335–1343. [DOI] [PMC free article] [PubMed] [Google Scholar]
  23. Calandruccio L, Porter HL, Leibold LJ, and Buss E (2020). “The Clear-Speech Benefit for School-Age Children: Speech-in-Noise and Speech-in-Speech Recognition,” J. Speech Lang. Hear. Res. 63, 4265–4276. [DOI] [PMC free article] [PubMed] [Google Scholar]
  24. Calandruccio L, and Smiljanić R (2012). “New sentence recognition materials developed using a basic non-native English lexicon,” J. Speech Lang. Hear. Res. 55, 1342–1355. [DOI] [PubMed] [Google Scholar]
  25. Cooke M (2006). “A glimpsing model of speech perception in noise,” J. Acoust. Soc. Am. 119, 1562–1573. [DOI] [PubMed] [Google Scholar]
  26. Cooke M, Garcia Lecumberri ML, and Barker J (2008). “The foreign language cocktail party problem: Energetic and informational masking effects in non-native speech perception,” J. Acoust. Soc. Am. 123, 414–427. [DOI] [PubMed] [Google Scholar]
  27. D’elattre P (1964). “Comparing the vocalic features of English, German, Spanish and French,” Internat Rev App Ling 2, 71–97. [Google Scholar]
  28. DePaolis RA, Janota CP, and Frank T (1996). “Frequency importance functions for words, sentences, and continuous discourse,” J. Speech Hear. Res. 39, 714–723. [DOI] [PubMed] [Google Scholar]
  29. Du Y, Shen Y, Wu X, and Chen J (2019). “The effect of speech material on the band importance function for Mandarin Chinese,” J. Acoust. Soc. Am. 146, 445–457. [DOI] [PMC free article] [PubMed] [Google Scholar]
  30. Eddins DA, Arnold M, Klein A, and Ellison J (2013). “Individual variability in unaided and aided measurement of the acceptable noise level,” Sem. Hear. 34, 118–127. [Google Scholar]
  31. Ezzatian P, Avivi M, and Schneider BA (2010). “Do nonnative listeners benefit as much as native listeners from spatial cues that release speech from masking?,” Speech Comm. 52, 919–929. [Google Scholar]
  32. Freyman RL, Balakrishnan U, and Helfer KS (2008). “Spatial release from masking with noise-vocoded speech,” J. Acoust. Soc. Am. 124, 1627–1637. [DOI] [PMC free article] [PubMed] [Google Scholar]
  33. Freyman RL, Helfer KS, and Balakrishnan U (2007). “Variability and uncertainty in masking by competing speech,” J. Acoust. Soc. Am. 121, 1040–1046. [DOI] [PubMed] [Google Scholar]
  34. Garcia Lecumberri ML, Cooke M, and Cutler A (2010). “Non-native speech perception in adverse conditions: A review,” Speech Comm. 52, 864–886. [Google Scholar]
  35. Golestani N, Rosen S, and Scott SK (2009). “Native-language benefit for understanding speech-in-noise: The contribution of semantics,” Biling. (Camb Engl) 12, 385–392. [DOI] [PMC free article] [PubMed] [Google Scholar]
  36. Hodgson M, Steininger G, and Razavi Z (2007). “Measurement and prediction of speech and noise levels and the Lombard effect in eating establishments,” J. Acoust. Soc. Am. 121, 2023–2033. [DOI] [PubMed] [Google Scholar]
  37. Iverson P, Kuhl PK, Akahane-Yamada R, Diesch E, Tohkura Y, Kettermann A, and Siebert C (2003). “A perceptual interference account of acquisition difficulties for non-native phonemes,” Cognition 87, B47–57. [DOI] [PubMed] [Google Scholar]
  38. Iyer N, Brungart DS, and Simpson BD (2010). “Effects of target-masker contextual similarity on the multimasker penalty in a three-talker diotic listening task,” J. Acoust. Soc. Am. 128, 2998–2910. [DOI] [PubMed] [Google Scholar]
  39. Jin IK, Lee J, Lee K, Kim J, Kim D, Sohn J, and Lee KJ (2016). “The Band-Importance Function for the Korean Standard Sentence Lists for Adults,” J. Audiol. Otol. 20, 80–84. [DOI] [PMC free article] [PubMed] [Google Scholar]
  40. Kilman L, Zekveld A, Hallgren M, and Ronnberg J (2014). “The influence of non-native language proficiency on speech perception performance,” Front. Psychol. 5, 651. [DOI] [PMC free article] [PubMed] [Google Scholar]
  41. Koelewijn T, Zekveld AA, Festen JM, Ronnberg J, and Kramer SE (2012). “Processing load induced by informational masking is related to linguistic abilities,” Int. J. Otolaryngol. 2012, 865731. [DOI] [PMC free article] [PubMed] [Google Scholar]
  42. Kuhl PK, Conboy BT, Coffey-Corina S, Padden D, Rivera-Gaxiola M, and Nelson T (2008). “Phonetic learning as a pathway to language: new data and native language magnet theory expanded (NLM-e),” Philos. Trans. R Soc. Lond. B Biol. Sci. 363, 979–1000. [DOI] [PMC free article] [PubMed] [Google Scholar]
  43. Kutlu E, Baxelbaum K, Sorensen E, Oleson J, and McMurray B (2024). “Linguistic diversity shapes flexible speech perception in school age children,” Sci. Rep. 14, 28825. [DOI] [PMC free article] [PubMed] [Google Scholar]
  44. Lenth RV (2025). “emmeans: Estimating Marginal Means, aka Least-Squares Means,” in R package version 1.11.0–003. [Google Scholar]
  45. Mayo LH, Florentine M, and Buus S (1997). “Age of second-language acquisition and perception of speech in noise,” J. Speech Lang. Hear. Res. 40, 686–693. [DOI] [PubMed] [Google Scholar]
  46. Moore BC, and Glasberg BR (1983). “Suggested formulae for calculating auditory-filter bandwidths and excitation patterns,” J. Acoust. Soc. Am. 74, 750–753. [DOI] [PubMed] [Google Scholar]
  47. O'Neill ER, Parke MN, Kreft HA, and Oxenham AJ (2020). “Development and Validation of Sentences Without Semantic Context to Complement the Basic English Lexicon Sentences,” J. Speech Lang. Hear. Res. 63, 3847–3854. [DOI] [PMC free article] [PubMed] [Google Scholar]
  48. R Core Team (2025). “R: A language and environment for statistical computing,” in R Foundation for Statistical Computing (Vienna, Austria: ). [Google Scholar]
  49. Recker KL, and Micheyl C (2017). “Speech Intelligibility as a Cue for Acceptable Noise Levels,” Ear Hear. 38, 465–474. [DOI] [PubMed] [Google Scholar]
  50. Regalado D, Kong J, Buss E, and Calandruccio L (2019). “Effects of Language History on Sentence Recognition in Noise or Two-Talker Speech: Monolingual, Early Bilingual, and Late Bilingual Speakers of English,” Am. J. Audiol. 28, 935–946. [DOI] [PMC free article] [PubMed] [Google Scholar]
  51. Shi LF, and Koenig LL (2016). “Relative Weighting of Semantic and Syntactic Cues in Native and Non-Native Listeners' Recognition of English Sentences,” Ear Hear. 37, 424–433. [DOI] [PubMed] [Google Scholar]
  52. Smits C, and Zekveld AA (2021). “Approaches to mathematical modeling of context effects in sentence recognition,” J. Acoust. Soc. Am. 149, 1371–1383. [DOI] [PubMed] [Google Scholar]
  53. Song J, and Iverson P (2018). “Listening effort during speech perception enhances auditory and lexical processing for non-native listeners and accents,” Cognition 179, 163–170. [DOI] [PubMed] [Google Scholar]
  54. Studebaker GA, and Sherbecoe RL (1991). “Frequency-importance and transfer functions for recorded CID W-22 word lists,” J. Speech Hear. Res. 34, 427–438. [DOI] [PubMed] [Google Scholar]
  55. Swaminathan J, Mason CR, Streeter TM, Best V, Kidd G Jr., and Patel AD (2015). “Musical training, individual differences and the cocktail party problem,” Sci. Rep. 5, 11628. [DOI] [PMC free article] [PubMed] [Google Scholar]
  56. Tamati TN, Gilbert JL, and Pisoni DB (2013). “Some factors underlying individual differences in speech recognition on PRESTO: a first report,” J. Am. Acad. Audiol. 24, 616–634. [DOI] [PMC free article] [PubMed] [Google Scholar]
  57. Tang C, van Heuven VJ, Heeringa W, and Gooskens C (2025). “Chinese “Dialects” and European “Languages”: A Comparison of Lexico‑Phonetic and Syntactic Distances,” Lang 10, 10060127. [Google Scholar]
  58. Wasiuk PA, Buss E, Oleson JJ, and Calandruccio L (2022). “Predicting speech-in-speech recognition: Short-term audibility, talker sex, and listener factors,” J. Acoust. Soc. Am. 152, 3010–3024. [DOI] [PubMed] [Google Scholar]
  59. Wong LL, Ho AH, Chua EW, and Soli SD (2007). “Development of the Cantonese speech intelligibility index,” J. Acoust. Soc. Am. 121, 2350–2361. [DOI] [PubMed] [Google Scholar]
  60. Yoho SE, Healy EW, Youngdahl CL, Barrett TS, and Apoux F (2018). “Speech-material and talker effects in speech band importance,” J. Acoust. Soc. Am. 143, 1417–1436. [DOI] [PMC free article] [PubMed] [Google Scholar]

RESOURCES