Ab initio word recognition in infant- and adult-directed continuous speech

Katie Von Holzen; Rochelle S. Newman

doi:10.1017/S0142716425100350

Ab initio word recognition in infant- and adult-directed continuous speech

Published online by Cambridge University Press: 29 December 2025

Katie Von Holzen

and

Rochelle S. Newman

Show author details

Katie Von Holzen*: Affiliation:
English and American Studies, Technische Universität Braunschweig, Braunschweig, Germany Department of Hearing and Speech Sciences, University of Maryland, College Park, MD, USA
Rochelle S. Newman: Affiliation:
Department of Hearing and Speech Sciences, University of Maryland, College Park, MD, USA
*: Corresponding author: Katie Von Holzen; Email: k.von-holzen@tu-braunschweig.de

Article contents

Abstract
Introduction
Experiment 1
Experiment 2
Experiment 3
General discussion
Conclusion
Supplementary material
Data availability statement
Financial support
Competing interests
Replication package
References

Rights & Permissions

Abstract

Continuous speech presents a challenge to the ab initio learner, as the language-specific segmentation strategies they use in their first language are not always reliable cues in other languages (Cutler 2001 International Journal of Research and Practice in Interpreting, 5(1), 1–23). Yet, they are able to use more general acoustic, prosodic, and statistical cues to word boundaries, as well as lexical similarity to their first language (e.g., Shoemaker & Rast 2013. Second Language Research, 29(2), 165–183) to recognize words at first exposure to a new language. In the current study, we investigated whether adult ab initio learners’ ability to recognize words after brief exposure to continuous speech in a new language is improved when that speech is produced using an infant-directed register, a style of speech found to facilitate segmentation in infancy (Thiessen et al. 2005. Infancy 7(1), 53–71). In a series of experiments, we demonstrate that although English ab initio learners of German benefited from infant-directed speech, their performance was generally lower than in previous studies unless task demands were reduced. These learners also benefited from word length and from frequency of occurrence, as has been shown previously, but these did not interact with register. As in infancy, learner-directed speech registers appear to facilitate initial processing and recognition in adult ab initio learners.

Keywords

Ab initio learners first exposure infant-directed speech word recognition

Information

Type: Original Article
Information: Applied Psycholinguistics , Volume 46 , 2025 , e49

DOI: https://doi.org/10.1017/S0142716425100350 [Opens in a new window]
Creative Commons: This is an Open Access article, distributed under the terms of the Creative Commons Attribution licence (https://creativecommons.org/licenses/by/4.0/), which permits unrestricted re-use, distribution and reproduction, provided the original article is properly cited.
Copyright: © The Author(s), 2025. Published by Cambridge University Press

Introduction

Many modern foreign language learning classrooms use immersive techniques to teach foreign languages, which include speaking exclusively in the to-be-learned language. In these immersive environments, it is likely that the teacher will not produce words in isolated instances with pauses in between, but instead use continuous utterances. Continuous speech presents a particular challenge to these beginning language learners, as the continuous input has to be segmented into individual words before meaning can be attached to the segmented sound strings. This has direct consequences for processing, as ab initio learners’ recognition of words in continuous speech is delayed and their electrophysiological responses reduced in comparison to native speakers (Snijders et al., Reference Snijders, Kooijman, Cutler and Hagoort2007). In the current study, we examine the conditions under which initial word recognition can be facilitated in adult ab initio learners, focusing on the prosodic modifications present in infant-directed speech.

Listeners’ approach to recognizing words in continuous speech is highly language-specific and is influenced by the language’s prosodic, phonological, and lexical structure (Cutler, Reference Cutler2001; Mattys & Bortfeld, Reference Mattys, Bortfeld, Gaskell and Mirkovic2016). Given the same input, native listeners of a stress-based language will segment it into words that match the input’s stress pattern, while native listeners of a syllable-based language will segment it into words that match the input’s syllable pattern (Cutler et al., Reference Cutler, Mehler, Norris and Segui1986; Mehler et al., Reference Mehler, Dommergues, Frauenfelder and Segui1981). The first language continues to exert an influence on strategies in subsequently acquired languages; even highly proficient L2 users continue to apply cues specific to their L1, even when these L1 cues are inappropriate for the context (Hanulíková et al., Reference Hanulíková, Mitterer and Mcqueen2011; Weber & Cutler, Reference Weber and Cutler2006; see also Finn & Hudson Kam, Reference Finn and Hudson Kam2008).

The ab initio learner ideally should therefore not rely on the same strategies that they use to recognize words in the speech of their native language. Although ab initio learners will continue to apply these incorrect L1 strategies, they have also been found to rely on more general, acoustic or prosodic cues to word boundaries. For example, words that occur near utterance boundaries are more readily recognized (e.g., Shoemaker & Rast, Reference Shoemaker and Rast2013), potentially due to a word boundary being marked by silence. The frequency of occurrence of a particular word structure in the native language may also facilitate its segmentation and recognition in a new language. Dutch ab initio learners more readily recognize bisyllabic compared with monosyllabic words from Mandarin Chinese (Gullberg et al., Reference Gullberg, Roberts and Dimroth2012), possibly because bisyllabic words are more frequent than monosyllabic words in Dutch (Vroomen et al., Reference Vroomen, Van Zon and De Gelder1996) or because bisyllabic words have a trochaic stress pattern in Mandarin (Lai et al., Reference Lai, Sui and Yuan2010) and initial stress is a cue for word boundaries in Dutch (e.g., Vroomen et al., Reference Vroomen, Van Zon and De Gelder1996).

Ab initio listeners begin to form lexical knowledge about a new language after only a brief exposure. Gullberg and colleagues (Gullberg et al., Reference Gullberg, Roberts and Dimroth2012) found that words that occurred more frequently in the input were more accurately recognized by novice listeners than infrequently occurring words. Frequency of occurrence in the input is thought to modulate the formation of a memory trace (Robinson, Reference Robinson2005). But beyond the frequency with which a form is encountered in the input, listeners are also sensitive to the statistical co-occurrence of syllables present in the input. They have been found to use the transitional probability between syllables, that is, the probability that syllable X will be followed by syllable Y or vice versa, to determine the boundaries between words after only a few minutes of exposure (Saffran et al., Reference Saffran, Newport, Aslin, Tunick and Barrueco1997). The evidence for the use of transitional probabilities has been demonstrated across the lifespan; adults, children, and infants as young as 5.5 months have been found to track the transitional probability between syllables in an artificial language (Johnson & Tyler, Reference Johnson and Tyler2010; Saffran et al., Reference Saffran, Aslin and Newport1996). Although many studies investigating the role of statistical learning in initial speech segmentation have used artificial languages, both adults (Alexander et al., Reference Alexander, Van Hedger and Batterink2023; Kittleson et al., Reference Kittleson, Aguilar, Tokerud, Plante and Asbjørnsen2010) and infants (Pelucchi et al., Reference Pelucchi, Hay and Saffran2009) have been found to track transitional probabilities in natural speech in a new language as well.

Thus, although initial word recognition provides a challenge to adult ab initio learners, they are able to exploit cues available in the input and form word-level representations after brief exposure. In the current study, we examine whether a type of input readily available for infant language learners similarly facilitates the subsequent formation of lexical knowledge in adult ab initio learners. Infant- or child-directed speech (IDS; CDS) describes the speech register used when adults are addressing infants and children. It differs from adult-directed speech (ADS) in acoustic characteristics, such as slower speech rate, higher and more varied pitch, and vowel hyperarticulation, as well as structural characteristics, such as simplified phrases and increased repetition (e.g. Fernald, Reference Fernald and Berkow1992; see Soderstrom, Reference Soderstrom2007 for a review). IDS has been proposed to support infants’ language acquisition by attracting infants’ attention to the speech signal (Rose et al., Reference Rose, Feldman and Jankowski2003) or through the modifications themselves, as infants’ word recognition is facilitated when words are presented in simple, repetitive word sequences, like those found in IDS (Fernald & Cummings, Reference Fernald and Cummings2003). The simplified phrasal structure and exaggerated prosody could also provide redundant grammatical cues, which would also support learning (e.g. Fisher & Tokura, Reference Fisher and Tokura1996; Morgan et al., Reference Morgan, Meier and Newport1987). Infants show enhanced word learning when the input is presented in IDS compared with ADS (Graf Estes & Hurley, Reference Graf Estes and Hurley2013; Ma et al., Reference Ma, Golinkoff, Houston and Hirsh-Pasek2011; Song et al., Reference Song, Demuth and Morgan2010; Zhou et al., Reference Zhou, Wang, Hong and Wong2023), suggesting that the use of IDS has important implications for language learning.

IDS may also be beneficial for speech segmentation and word recognition. IDS often presents longer pauses at phrasal boundaries than ADS (Fernald & Mazzie, Reference Fernald and Mazzie1991) and phrases produced in IDS have on average fewer words than those produced in ADS, which can result in a slower speech rate, a greater percentage of words in the utterances already having one word boundary marked, and more opportunities for phrase-final lengthening (Martin et al., Reference Martin, Igarashi, Jincho and Mazuka2016). Although not universally present, such features could aid segmentation (e.g., Gout et al., Reference Gout, Christophe and Morgan2004; White et al., Reference White, Benavides-Varela and Mády2020). Pitch peaks are also higher in IDS in comparison to ADS (Fernald et al., Reference Fernald, Taeschner, Dunn, Papousek, De Boysson-Bardies and Fukui1989). Pitch peaks are a reliable indicator of stress (Jessen et al., Reference Jessen, Marasek, Schneider and Clahßen1995; Sluijter & Van Heuven, Reference Sluijter and Van Heuven1996), which in turn serves as a word boundary cue in stress-based languages (Cutler, Reference Cutler1994) and both English- and German-learning infants successfully use the dominant stress pattern, the trochaic unit, to segment and recognize words (Höhle et al., Reference Höhle, Bijeljac-Babic, Herold, Weissenborn and Nazzi2009; Jusczyk et al., Reference Jusczyk, Houston and Newsome1999). However direct comparisons of the influence of IDS and ADS speech on segmentation ability in infants have yielded mixed results. Singh and colleagues (Singh et al., Reference Singh, Nestor, Parikh and Yull2009) found that previous exposure to words produced in English IDS, but not ADS, led to successful word recognition 24 hours later. In contrast, infants learning a language other than English, German, are equally successful at recognizing words presented in IDS and ADS, although this may be due to the language-specific characteristics of German IDS (Mani & Paetzold, Reference Mani and Paetzold2016; Schreiner et al., Reference Schreiner, Altvater-Mackensen and Mani2016).

Infants can also more easily exploit the statistical cues present in speech when it is produced using an IDS as opposed to an ADS register, supporting word segmentation. Thiessen et al. (Reference Thiessen, Hill and Saffran2005) familiarized English-learning 7-month-olds to an artificial language. Half of the infants heard the language produced with an IDS register, while the other half heard it produced with an ADS register, with recordings modified to have the same overall amplitude and length as well as the occurrence of within-word pitch accents. This ensured that the differences between the IDS and ADS stimuli were due to their pitch contour and fundamental frequency (F0). Infants listened longer to words compared to part words only when they were familiarized with IDS, suggesting that the exaggerated pitch and intonation in this register support speech segmentation.

IDS may be beneficial for adult ab initio learners. In a series of experiments, Golinkoff and Alioto (Reference Golinkoff and Alioto1995) familiarized English-speaking adults with pictures paired with sentences in Chinese labeling the picture (Zhe shi yi ge ping—This is a bottle). Performance at word learning was higher than chance only when participants were familiarized with IDS, but not ADS speech. The word-learning studies of Golinkoff and Alioto (Reference Golinkoff and Alioto1995) and a similar study of Ma et al. (Reference Ma, Fiveash, Margulis, Behrend and Thompson2020) require the learner to first segment the words from the familiarization sentences and then map the word onto the referential object they were viewing at the time. Although these results suggest that IDS is helpful, it is not clear whether this advantage occurs during segmentation, label-referent association, or both. Indeed, Ma and colleagues found that familiarizing participants with the targets produced in IDS in isolation led to better subsequent word recognition than targets produced in ADS, suggesting that facilitation is at least happening during label-referent association. It therefore remains unclear to what extent IDS aids adult learners in word recognition when the component of label-referent association is removed.

In the current study, we examine whether the exaggerated pitch changes found in infant-directed speech influence adult ab initio learners’ ability to recognize words they have been exposed to in a new, natural language. Using a modification of the paradigm used by Gullberg et al. (Reference Gullberg, Roberts and Dimroth2012), we compared English-speaking adults’ ability to recognize word forms that were previously presented in continuous speech produced using an IDS or ADS register (target items). Although IDS is typically produced more slowly than ADS, we controlled for this rate difference, such that speaking rate/duration differences were unable to impact recognition. To be more compatible with previous literature, we analyze participants’ accuracy in the form of word acceptance, or their likelihood to indicate that they previously heard a test item during exposure (Gullberg et al., Reference Gullberg, Roberts and Dimroth2012; Kittleson et al., Reference Kittleson, Aguilar, Tokerud, Plante and Asbjørnsen2010). We also analyzed participants’ d’ scores to capture their sensitivity to the presence of target items (Hautus et al., Reference Hautus, Macmillan and Creelman2021). We predicted that if IDS facilitates word recognition in ab initio learners, then in comparison to participants exposed to sentences produced with ADS, participants exposed to sentences produced with IDS would show (1) both word acceptance scores for target items and d’ scores above chance levels (50% and 0, respectively), (2) greater acceptance rates for target compared to filler items, and (3) greater d’ scores.

We also examined the use of both word length and frequency. If ab initio listeners more readily recognize target items with a word length more common in their native language, in our case monosyllabic words (Cutler & Carter, Reference Cutler and Carter1987), then we expect that in comparison to bisyllabic items, responses to monosyllabic items will show a greater difference in word acceptance rates between target and filler items as well as greater d’ scores. Alternatively, if an advantage for monosyllabic over bisyllabic target items is found, it could be due to the word length effect (Baddeley et al., Reference Baddeley, Thomson and Buchanan1975), as the shorter monosyllabic items are easier to hold in short-term memory. Finally, if frequency supports the formation of a memory trace for items in the input (Gullberg et al., Reference Gullberg, Roberts and Dimroth2012; Robinson, Reference Robinson2005), then frequently presented target items should show higher acceptance rates than infrequently presented target items when compared with chance items.

The present article was written with embedded analysis scripts in R (R Core Team, 2018) using the papaja package (Aust & Barth, Reference Aust and Barth2018) in R Markdown (Allaire et al., Reference Allaire, Xie, McPherson, Luraschi, Ushey, Atkins and Chang2018). The R Markdown and R analysis scripts as well as the de-identified data are available on Open Science Framework (https://osf.io/f2mqe/?view_only=25317ee805684313bb651dcabb37fb31). The stimuli and programs to run all three experiments are also included.

Experiment 1

Methods

Participants

Participants were recruited from the University of Maryland—College Park. Half of the participants (n = 21; M age = 20.81 years, SD age = 3.27; range 18–34, 13 women) were exposed to sentences produced with infant-directed speech (IDS) and the other half of the participants were exposed to sentences produced with adult-directed speech (ADS; n = 22; M age = 20.64 years, SD age = 1.92; range 18–26, 13 women). These sample sizes were chosen for their similarity to the sample sizes of (Gullberg et al., Reference Gullberg, Roberts and Dimroth2012), which had a similar experimental design in terms of between- and within-subjects manipulations. Before the experiment, each participant filled out a language use and background questionnaire to ensure that American English was the participant’s native language and that they had not had previous exposure to German or another West Germanic language besides English (e.g. Dutch, Afrikaans). All participants in the final sample had normal hearing and normal or corrected vision. An additional 6 participants were tested but not included in the final sample because they didn’t complete the experiment (n = 1), reported uncorrected vision impairment (n = 1), were not a native speaker of English (n = 2), or had learned German (n = 2). All participants were paid a small amount of money for their participation.

Stimuli

Following the design of Gullberg and colleagues (Gullberg et al., Reference Gullberg, Roberts and Dimroth2012), participants heard sentences during the exposure phase, but individual words during the test phase. Ninety-six real German words were selected to serve as target (n = 24; 12 monosyllabic, 12 bisyllabic) and filler (n = 72; 36 monosyllabic, 36 bisyllabic) items for the test phase of this study (see Appendix A1 for a list of all target and filler items). All bisyllabic items were stress-initial. No syllable was used more than once across test items, ensuring that the target and filler items were maximally different from one another. Although these target and filler items were themselves phonotactically legal in English and contained sounds with phonetic features similar to English phonemes, the sentences in which the target items were presented contained some non-English phonetic features, such as the vowels /ʏ/, /o:/, /e:/, and /a/ and the affricates /ts/ and /pf/.

The 24 target items were used to create sentences for the exposure phase. For each target item, a set of eight grammatically correct and semantically meaningful standard German sentences were created (see Appendix A2 for sentence lists), resulting in a total of 192 sentences. The target item was never placed in the initial or final position within the sentences, as utterance boundaries may facilitate recognition (Shoemaker & Rast, Reference Shoemaker and Rast2013). To control for arbitrary listening preferences and modulate the frequency of occurrence of target items during the exposure phase, two sets of sentences using a different combination of 120 sentences were created (training sets A and B). For example, in training set A, half of the target items (6 monosyllabic, 6 bisyllabic) appeared frequently (8 times each, making up 96 of the 120 trials in the exposure phase) while the other half of the target items (6 monosyllabic, 6 bisyllabic) appeared infrequently (2 times each, for 24 of the 120 trials in the exposure phase). Training set B had the same structure, but the frequent and infrequent target items were switched. For example, if Rindvieh was a frequent item in training set A, appearing 8 times, it would be an infrequent item in training set B, appearing 2 times. Each sentence had on average 10.41 syllables (SD = 2.05; range 5–15). Neither filler items nor their constituting syllables ever occurred in the sentences. Whether a participant heard training set A or B was counterbalanced across participants and registers (IDS/ADS).

Artificial language studies often use a handful of syllables in combination with one another to form words while controlling for their statistical probability. The current study used a natural language, with 441 and 445 unique syllables occurring in training sets A and B respectively. To ensure that the syllables making up the target items did not vary drastically in their transitional probability (TP) in the training sets, the TP of target item syllables was controlled for. For bisyllabic target items, the internal TP was always 1.0. The first syllable was always followed by the second syllable, and both syllables never occurred anywhere else in the sentences. For frequent mono- and bi-syllabic target items, the forward and backward TPs were 0.25 or less. This means that frequent target items could be preceded or followed by the same word no more than two times across sentences. For infrequent mono- and bi-syllabic target items, the forward and backward TPs were 0.5. Infrequent target items only appeared twice and were always preceded and followed by different words across sentences. A summary of the TPs of the target items in both training sets can be found in Appendix B.

A female, native speaker of German produced the training sentences several times in random order in both IDS and ADS. Due to the nature of IDS and ADS, the IDS recordings were longer and louder than the ADS. Similar to Thiessen et al. (Reference Thiessen, Hill and Saffran2005), we addressed these possible confounds by adjusting the length and amplitude of the sentences. For each sentence, we calculated the duration of the IDS and ADS versions as well as the average of the two. Next, the duration of the IDS and ADS versions of the sentence were adjusted using the “stretch” algorithm in Adobe Audition (Adobe System Corporation, San Jose, CA) to this average. The resulting IDS and ADS sentences were therefore matched in duration (M = 2.83 ms; SD = 0.47) and speech rate (M = 3.68 syllables/sec; SD = 3.68) for the training sentences. Note that this controls for overall speech rate, but does not adjust for the exaggerated sentence- and word-final lengthening of IDS (Albin & Echols, Reference Albin and Echols1996). Using Praat (Boersma & Weenink, Reference Boersma and Weenink2016), the training sentences produced in both IDS and ADS were adjusted to the same average intensity in dB SPL. In addition, the same speaker recorded each of the 24 target items as well as 72 filler items in isolation for a total of 96 test items. These recordings were also adjusted to the same average intensity level in dB SPL.

The speaker’s average F0 was measured for each sentence. Average F0 in Hertz was significantly higher for the IDS sentences (264.86 Hz, SD = 20.94, 338.98 mel) compared with the ADS sentences (171.15 Hz, SD = 23.14, 227.93 mel), t(191) = 36.69, p < .001, d = 2.65. The speaker’s average minimum and maximum F0 values in Hertz for the IDS sentences (Minimum: 137.58 Hz, 185.97 mel; Maximum: 525.83 Hz, 609.59 mel) were also significantly higher than the average minimum and maximum F0 values for the ADS sentences (Minimum: 94.26 Hz, 129.96 mel; Maximum: 352.64 Hz, 435.78 mel), respectively, t(191) = 16.94, p < .001, d = 1.22 and t(191) = 14.05, p < .001, d = 1.01. These values are consistent with F0 characteristics of IDS and ADS speech in German, as reported by previous studies (Fernald et al., Reference Fernald, Taeschner, Dunn, Papousek, De Boysson-Bardies and Fukui1989; Schreiner et al., Reference Schreiner, Altvater-Mackensen and Mani2016). For the target items produced in isolation, the speaker was instructed to use a register between that of IDS and ADS, in order to reduce the likelihood that a match in register between the stimuli used at exposure and test could explain any differences in responses between the IDS and ADS registers. The speakers’ average F0 for the target items was 239.31 Hz (SD = 49.50, range average = 157.76–361.80; M mel = 309.54, range = 211.34–445.54 mel). Regarding pitch values, the target items in the test phase were more similar to the IDS exposure sentences in average and minimum F0, but more similar to the ADS exposure sentences in maximum F0.

Design and procedure

The experiment was presented on a 2017 iMac computer using the Open Sesame software (Mathôt et al., Reference Mathôt, Schreij and Theeuwes2012). The experiment began with participants seated in front of a computer wearing headphones. They were instructed to simply listen to the sentences. Then, the exposure phase began and the 120 sentences were presented randomly, with the constraint that sentences containing the same target item could not occur within 5 sentences of one another. Presentation of the exposure phase lasted about 6 minutes.

After completing the exposure phase, instructions appeared on the screen, telling participants that they would be played an audio file and then would be prompted to indicate whether they had heard the sound in the exposure phase or not by pressing the corresponding key (key assignment was counterbalanced across participants). After these instructions, the test phase began. Target and filler items were presented randomly, one time each, with the constraint that items of each type (target/filler) could not be presented more than 2 times in a row. Once participants indicated their response, a fixation cross appeared on the screen for 1,000 ms, indicating that the trial had been completed. The correct expected response for all 24 target items was “present” for both training sets A and B and the correct expected response for all 72 filler items was “not-present”.

Data analysis

Participant responses were analyzed using R (R Core Team, 2018). Analyses of word acceptance (coded as 1 for a “present” response and 0 for a “not present” response) were modeled using mixed-effect logistic regression models (Bates et al., Reference Bates, Maechler, Bolker and Walker2015). We used the “order” command in the buildmer package (Voeten, Reference Voeten2023) to determine the maximal random effects structure, including random slopes for the random intercepts of participant and item. All contrasts for fixed effects were deviation coded (−1, 1), with the exception of Frequency Status (see below). The emmeans package (Lenth, Reference Lenth2019) was used to report estimated marginal means for significant effects and interactions and to test post-hoc comparisons. An initial model was conducted with the intercept removed to compare target item performance against chance.

We examined the influence of Frequency Status (Frequent Target, Infrequent Target, Filler), Register (IDS, ADS), and Word Length (monosyllabic, bisyllabic) on participants’ word acceptance scores. The factor of Frequency Status reflected target items that had appeared frequently (Frequent Target) and infrequently (Infrequent Target) in the exposure phase, as well as filler items that did not appear in the exposure phase. It was deviation coded such that frequently and infrequently presented target items were separately compared to filler items. In the case of an effect of or interaction with Frequency Status, differences in word acceptance responses to frequently and infrequently presented targets were established using post-hoc comparisons of the estimated marginal means.

Analyses of d’ examined participants’ sensitivity to discriminate between target and filler items based on Signal Detection Theory (Stanislaw & Todorov, Reference Stanislaw and Todorov1999). For each participant, we calculated the number of hits (correctly identifying a target item as “present”), correct rejections (correctly identifying a filler item as “not present”), misses (incorrectly identifying a target item as “not present”), and false alarms (incorrectly identifying a filler item as “present”). We used the psych package (Revelle, Reference Revelle2023) in R (R Core Team, 2018) to calculate d’ separately for frequently and infrequently presented target items, but using the same filler items for correct rejections and false alarms. Analyses of d’ were modeled using mixed-effect linear regression models (Bates et al., Reference Bates, Maechler, Bolker and Walker2015). As in the analysis of word acceptance, we used the “order” command in the buildmer package (Voeten, Reference Voeten2023) to determine the maximal random effects structure, with the exception that no random intercepts for item were included as these scores cannot be calculated at the item level. All contrasts for fixed effects were deviation coded (−1, 1) and the emmeans package (Lenth, Reference Lenth2019) was used to report estimated marginal means for significant effects and interactions and to test post-hoc comparisons. An initial model was conducted with the intercept removed to compare d’ against chance.

To ensure responses did not differ between the two training sets, we compared word acceptance for target and filler items and d’ scores between the two training sets (training sets A, B). A general mixed-effects model of word acceptance with an interaction between Frequency Status (Frequent Target, Infrequent Target, Filler) and Training Set (A, B) had a maximum random effects structure of random intercepts for subjects and items. The main effect of Training Set (β = 0.01, SE = 0.1, Z value = 0.07, p = 0.95) and the interaction between Training Set and Frequency status (Frequent Targets: β = 0.01, SE = 0.13, Z value = 0.08, p = 0.93; Infrequent Targets: β = 0.02, SE = 0.13, Z value = 0.13, p = 0.9) were not significant. A t-test comparing d’ for participants exposed to the two training sets (A and B) revealed no significant difference, t(40.91) = 0.41, p = .682. Subsequent analyses were collapsed across training set.

Results

Analyses of word acceptance

Table 1 and Figure 1 display the descriptive statistics for word acceptance for the factors of Frequency Status (Frequent Target, Infrequent Target, Filler), Register (IDS, ADS), and Word Length (monosyllabic, bisyllabic). For the comparisons of target item word acceptance to chance, the full output of the model without an intercept is given in Appendix C1. The maximum random effects structure had random intercepts for subjects and items, with a random slope for Word Length on the subjects intercept and Register on the items intercept. Word acceptance was significantly above chance for frequent, bisyllabic target items when participants were exposed to sentences produced with IDS (β = 0.63, SE = 0.28, Z value = 2.25, p = 0.02). All other comparisons to chance were not significant.

Table 1.

Means, standard deviations, and 95% confidence intervals for word acceptance rates for monosyllabic and bisyllabic target items, presented frequently and infrequently, as well as filler items for participants exposed to sentences produced with IDS or ADS in Experiment 1

Figure 1.

Mean word acceptance scores (%) with +/−1 standard error in Experiment 1 for frequently and infrequently presented target items and filler items. Responses to monosyllabic and bisyllabic test items are given in the top and bottom panels, respectively. Responses for participants trained in IDS are given in the left panels, while responses for those trained in ADS are given in the right panels. The red line at 50% indicates the chance level.

Analyses of d’

Table 2 displays the descriptive statistics for d’ for the factors of Word Length (monosyllabic, bisyllabic), Frequency (frequent, infrequent), and Register (IDS, ADS). As can be seen in Table 2, the values of beta were around 1.0, an indication that participants were not biased overall in their responses to accept or reject items (Hautus et al., Reference Hautus, Macmillan and Creelman2021). For the comparisons of d’ to chance, the full output of the linear mixed-effects model without an intercept is given in Appendix C3. The maximum random effects structure had a random intercept for subjects, with a random slope for Word Length. D’ scores were significantly above chance for frequent, bisyllabic target items when participants were exposed to sentences produced with IDS (β = 0.27, SE = 0.11, t value = 2.45, p = 0.02). All other comparisons to chance were not significant.

Table 2.

Means, standard deviations, and 95% confidence intervals for d’ and beta for bisyllabic and monosyllabic items presented frequently and infrequently to participants exposed to sentences produced with IDS or ADS in Experiment 1

Discussion

The results of Experiment 1 revealed evidence that L1 American-English speakers without any prior exposure to German were able to recognize words after only brief exposure to this language. Yet this was clearly a difficult task. Perhaps as a result, this finding was restricted to target items heard frequently (8 times) during the exposure phase, and strongest for bisyllabic items and in participants that were exposed to sentences produced with IDS during the exposure phase. This tentatively confirms our predictions that IDS and frequent presentation of target items during exposure aids initial word recognition in ab initio learners. Evidence of recognition for bisyllabic but not monosyllabic items runs counter to our predictions that ab initio learners should more readily segment words with a word length more common in their native language.

Strong conclusions about recognition in Experiment 1 should be interpreted with caution. Although the analysis of word acceptance suggests that participants reliably differentiated between filler items and target items that had been presented previously in the exposure phase, d’ scores were relatively close to 0, suggesting that participants had difficulty discriminating between target and filler items, regardless of whether they were trained using IDS or ADS or whether those items were monosyllabic or bisyllabic. Thus, while the limited effect found in the chance data is suggestive, more evidence is needed to accept these results fully.

The paradigm used in Experiment 1 required participants to listen to speech in an unknown language and hold memory traces of heard syllable strings for several minutes. Without additional cues, such as visual aids and gestural highlighting used by Gullberg et al. (Reference Gullberg, Roberts and Dimroth2012), this may tax listeners’ working memory abilities. In Experiment 2, we examine whether ab initio learners are able to recognize words previously presented in our training sentences when the burden on their working memory is removed. We use the paradigm employed by Shoemaker and Rast (Reference Shoemaker and Rast2013) where presentation of a sentence is immediately followed by presentation of a test item that either was or was not in the preceding sentence. In this paradigm, participants are immediately prompted to determine whether the test item was present in the sentence or not. If participants’ low performance in Experiment 1 was due to the sentence stimuli, either the stimuli items used or the recordings themselves, we should not find successful recognition when testing it using a different paradigm, and should find similar low performance using a different paradigm in Experiment 2. If their low performance is the result of the high memory demands, we should find higher performance overall in Experiment 2, but the same pattern of results. Such a replication of the basic pattern would strengthen the arguments from this current study.