1.1. Papuan Malay stress in production and perception

An acoustic study investigated several temporal, spectral and amplitudinal cues to word stress in spontaneously produced Papuan Malay monologues from 19 speakers (Kaland, 2019). The analysed words were all disyllabic, which matches the most common word length in this language, and were taken from nonfinal positions in the intonation unit (IU). It has been shown that IU-final words exhibit rather exceptional prosodic features, i.e., the largest f0 movements, final lengthening, etc. (Kaland & Baumann, 2020). Results showed that duration per phoneme (syllable duration divided by number of phonemes) was the strongest acoustic correlate of word stress (stressed = longer), followed by vowel displacement (stressed = more peripheral), and spectral tilt (stressed = shallower intensity roll-off towards higher frequencies). All these cues also showed larger acoustic differences between stressed and unstressed syllables for ultimate stress compared to penultimate stress.

The role of the acoustic cues in word recognition was tested in a series of experiments in which listeners guessed which word matched with the hummed and acoustically manipulated syllable sequence they were hearing (Kaland, 2020). The acoustic manipulations rendered the stress pattern in the disyllabic sequences either as penultimate or ultimate using the acoustic values for each of the cues as found in production (Kaland, 2019). The participants’ task was to match the manipulated sequences with the correct stress pattern of the word. Results showed higher correctness scores were obtained for patterns presented in isolation than for those in a phrase, and higher ones for ultimate stress than for penultimate stress. The latter result reflected the stronger acoustic marking of ultimate stress compared to penultimate stress in the production analysis in Kaland (2019). However, scores were overall just above chance level, indicating a minimal benefit from the cues when tested separately (53%–60% correct).

Two lexical analyses (Kaland et al., 2021) investigated Papuan Malay word stress on the basis of a representative word list of native roots (a “lexicon” of approximately 1000 words from Kluge, 2017). First, it was investigated whether stress information could help to disambiguate between words. This was done by counting word embeddings, i.e., words that are embedded in longer ones. For example in English, bee is an embedding in words such as belay and beanie. When stress information is taken into account, bee (stressed) is only an embedding in beanie, matching the initial stressed syllable, and no longer in belay, not matching the initial unstressed syllable (e.g., Cutler et al., 2004; Cutler & Pasveer, 2006). The amount of reduction in word embeddings as a result of taking stress information into account is an indication of the extent to which stress information can facilitate word recognition. Thus, listeners could discard potential word candidates when taking stress information into account, with more reduction in the number of embeddings indicating a larger degree of facilitation. It was found that in Papuan Malay, the number of embeddings could be reduced to the same extent as in English and to a lesser extent than in Spanish, Dutch, or German. In a second lexical analysis, the phonological factors determining the position of stress in the word were investigated. This was done by comparing more than 20 variables relating to the segmental and syllabic makeup of the words. It was confirmed that /ɛ/ in the penultimate syllable, which reduces to schwa at the surface, is the main reason for stress to shift to the ultimate position. Additionally /ɔ/ was found to reject stress in the ultimate syllable as well (i.e., no shift from penultimate to ultimate). Taken together, these two phonological rules could be subsumed under the rejection of stress by mid vowels (/ɛ/ and /ɔ/ being the only mid vowels in Papuan Malay; Kluge, 2017), together explaining more than 96% of the stress patterns in the wordlist.

In order to investigate the extent to which Papuan Malay listeners could actually make functional use of the word stress cues in word recognition, a gating task was carried out (Kaland, 2021). Gating tasks have been used in word recognition experiments by presenting word parts of increasing length to listeners (e.g., Cotton & Grosjean, 1984). For Papuan Malay for example, the syllable /bɛ/ was presented to listeners in a task where they had to indicate whether that syllable was taken from ['bɛ.bɛk], in which it was stressed, or from [bɛ.'ban], in which it was unstressed. As control gates, listeners were presented either with a single phoneme [b] (chance level responses expected) or with the entire word (100% correct expected). The task presented the target word written on the screen and listeners indicated which out of two auditory gates matched the target word. Listeners thus heard a gate from a stressed syllable and from an unstressed one before giving their response. Results showed that listeners could correctly identify the target word in more than 80% of the cases based only on the syllable, indicating that they had successfully used the acoustic cues to stress in this task (i.e., well above chance level).

A crucial remaining question from the latter study is whether listeners were successful because they made use of lexically stored stress information, or whether they could use the acoustic contrast between the two presented gates to match what they heard with the target on the screen. The latter option does not necessarily demand lexically stored stress information, as listeners could retrieve the crucial cues from their auditory working memory, having heard those cues in the experiment. In other words, the gating task did not challenge listeners’ memory of stress information. Notably, participants also did not score 100% correct, as in the control gate presenting the entire word. This indicates that stress information helped listeners, but not to the extent that is expected when that information is unambiguously coupled to the word. It is therefore not entirely clear to what extent Papuan Malay stress patterns are stored lexically. As further outlined in the next section, the idea is that highly predictable patterns do not demand lexical storage. The question is therefore whether Papuan Malay listeners are able to store stress information in a task in which their memory is directly tested.

1.2. Testing lexical stress storage in sequence recall tasks

Discrimination tasks were carried out in which listeners were presented with three auditory stimuli and had the task to decide whether the third stimulus was more similar to the first or the second (henceforth ABX discrimination tasks). These showed that French listeners made many more errors distinguishing nonwords that only differ in the position of stress compared to Spanish listeners (Dupoux et al., 1997). The difference was attributed to whether word stress has a contrastive function in these languages. In French, the position is fixed on the final syllable, whereas in Spanish, it is highly variable, giving rise to minimal stress pairs that can only be successfully disambiguated when listeners are able to use their acoustic cues. Note that there are different analyses of the final (word) prominence in French, i.e., it has “fixed stress” (Di Cristo, 1998, p. 196), “no primary word stress at all” (Van der Hulst, 2012, p. 1515), it is a language for which “the need for w-final accents is not hard to establish” (Gussenhoven, 2004, p. 258), or “the lexical representation of a word does not include metrical or tonal properties” (Delais-Roussarie et al., 2015, p. 65). French listeners were reported to exhibit stress deafness, i.e., not being literally deaf to the acoustic cues, but unable to use them in word recognition (Dupoux et al., 1997). In the remainder of this paper I therefore refrain from using this term.

The Spanish-French difference in performance in the ABX task was hypothesized to originate from differences in the ability to memorize varying stress patterns. The French listeners were expected to not have such experience from their every day use of their language, whereas Spanish listeners would need to do so all the time. There was, however, considerable overlap in the performance of participants, with some French participants being as successful as Spanish ones and some Spanish participants showing as much difficulty as French ones (Dupoux et al., 1997). A different paradigm was developed in order to test the hypothesized differences more directly in stress memory (Dupoux et al., 2001). This paradigm concerned a recall task of nonce word sequences that differed in phonemes (e.g., ['ku.pi] or ['ku.ti]) or in stress (e.g., ['mi.pa] or [mi.'pa]). Sequences consisted of two to six nonce words, and listeners showed overall more recall difficulty as sequences got longer. For each subsequent experiment in Dupoux et al. (2001) more variation in the stimulus material was added, i.e., synthesized pitch variation in addition to the other acoustic cues, variation in the vowel makeup of the nonce words, and the addition of another speaker producing the nonce words. With each additional type of variation, French listeners showed worse recall performance for the stimuli that had stress differences, whereas the Spanish listeners’ performance was stable throughout the series of experiments. In a final experiment, all acoustic variation was taken away by testing French and Spanish participants using a single token from the stimulus material. The results showed similar performance for both language groups, in that French listeners performed clearly better compared to the previous experiments that had more sources of acoustic variation.

The study explained the results by three perceptual strategies to store incoming speech in the auditory working memory (Dupoux et al., 2001). The first strategy concerns explicit categorization by matching the auditory input with stored representations in the lexicon, which gets more difficult with longer sequences. The second is a mismatch detection mechanism between two recent subsequent units, i.e., whether they were the same or not. The second strategy worked well for the French listeners when the nonce words only differed for the crucial experimental cues (i.e., segmental/stress) and had no other naturalistic acoustic variation (i.e., pitch, vowels, speaker’s voice). The third strategy concerns the ability to store phonological representations of stress patterns, which is only available for listeners of a stress language (i.e., Spanish).

Phoneme/stress recall tasks (henceforth PSRTs) involving multiple sources of acoustic variation have been reported in several studies since Dupoux et al. (2001). In what follows, the main findings of these studies are summarized and a schematic overview of them is given in Table 1. The accuracy in each of the PSRTs is expressed using an index proposed in Peperkamp & Dupoux (2002), which subtracts the error rate (percentage) for phoneme contrasts from the error rate for stress contrasts. A critical discussion on the cross-study comparability is given in Section 1.3, providing a summary of the main findings of all discussed studies in Table 1.

1.3. Comparability of PSRTs

Note that the reported measures of recall accuracy differed among the discussed studies (Table 1). The index proposed in Peperkamp & Dupoux (2002) is based on error rates, which can be straightforwardly obtained from correctness rates (i.e., as percentage: error rate = 100 – correct rate). The index is the result of subtracting phoneme error rates from stress error rates, as a measure of listeners’ performance (index = stress error rate – phoneme error rate). Thus, the index takes the errors made for phoneme contrasts as a baseline to express the degree of difficulty listeners had with stress contrasts. This measure thus accounts for baseline differences in the perception of phoneme contrasts across languages. It is in itself intriguing that in the typological studies, using highly comparable PSRT paradigms, baseline (phonemic) error rates differed among some languages (Peperkamp & Dupoux, 2002; Peperkamp et al., 2010; Rahmani et al., 2015). The index thus abstracts over number of items and conditions, and as such seems to be suitable to compare recall accuracy across studies. Note that the index, when showing negative values, also accurately captures that recall performance was better for stress contrasts than for phoneme contrasts (Dupoux et al., 2001; Peperkamp & Dupoux, 2002; Dupoux et al., 2008; Heisterueber et al., 2014). However, the index does not reflect overall differences among listeners, i.e., due to language impairment (Haake et al., 2013). That is, children with poor recall performance for both segmental and stress contrasts show a similar index value compared to the typically developing ones (Table 1).

It is also important to note that the studies on Tashlhiyt Berber, Moroccan Arabic, and Maltese (Bruggeman, 2020; Lialiou et al., 2023) concluded that listeners in those languages pattern with the lowest performing ones in Rahmani et al. (2015). However, their index values are in fact similar to the best performing listener groups (Dutch and Japanese) in Rahmani et al. This is a puzzling observation given that identical stimuli and procedures were used across all three studies. Note that none of the three studies reported index values directly, i.e., they applied mixed effects modelling on the correctness rates with, among others, contrast (phoneme/stress) as a factor. It is beyond the scope of this study to apply this type of modelling on the older studies in Table 1. For this reason, index values (computed post hoc if not explicitly reported) are chosen to express the results using the same metric for all studies, as they allow for (limited) comparison of the PSRT studies.

The index values suggest that Tashlhiyt Berber, Moroccan Arabic, and Maltese listeners perform better than the French listeners in earlier studies (e.g., Dupoux et al., 2001; Peperkamp & Dupoux, 2002; Dupoux et al., 2008; Peperkamp et al., 2010). Given some controversial reports in the literature on the status of word stress in Tashlhiyt Berber, Moroccan Arabic, and Maltese (e.g., Gordon & Nafi, 2012; Boudlal, 2001; Azzopardi, 1981), the index values in Table 1 might be open to different interpretations, i.e., falling somewhat in between languages that maximally rely on lexically stored stress information (e.g., Spanish) and languages that do not at all (e.g., French). At the same time, French listeners in Peperkamp & Brazeal (2023) show lower index values than in any of the previous studies involving French participants. Given the high comparability of the methods applied to French participants, it is reasonable to assume that there are multiple sources of variation affecting the index values across studies. Within-study comparisons of the index values are likely to be more reliable than across-study comparisons.

Thus, the studies discussed in the previous sections are comparable to a limited extent, although they all made use of a paradigm that tested both phonemic and stress contrasts using recall tasks. There are potentially multiple aspects of the studies that limit their comparability. For example, stimulus materials were generally adapted to the language in order to guarantee that they did not match existing words or to test specific additional prosodic aspects such as type of stress pattern (Haake et al., 2013; Heisterueber et al., 2014) and phrase position (Correia et al., 2015). Furthermore, studies differed in their experimental setup in that the number and type of training rounds prior to the PSRT varied (cf. Dupoux et al., 2001, Heisterueber et al., 2014, Rahmani et al., 2015). Sequence length is likely another source of variation affecting the index values in Table 1. It is, however, unclear how exactly sequence length affects the average index values. One could hypothesize that with longer sequences (e.g., 5 or 6 nonce words), language differences come out stronger in the index values as listeners have higher error rates for longer sequences (i.e., large range of index values) compared to shorter sequences (e.g., ≤4 nonce words), but this pattern does not hold across the studies (see Table 1). By averaging over sequence lengths, the index value does not capture the steepness of the error rate increase when sequences get longer. This rate might also vary across languages, as can be observed in the results reported in Rahmani et al. (2015). Another source of variation in the index values might come from phrase prosodic expectations of the listeners. These vary largely between the studied languages, even more so when taking into account how they might have interacted with word prosodic ones. The PSRT involves an item familiarisation phase and, given the length of the items, they are likely stored as lexical items comparable to words. However, presented in a sequence and with varying prosodic patterns, listeners’ recall performance could have been influenced by the prosodic expectations of larger units (i.e., phrases). In fact, phrase prosodic context (nuclear, postnuclear) was shown to affect European Portuguese listeners’ recall scores in a PSRT (Correia et al., 2015). Furthermore, some studies reported a high level of individual variation across listeners’ performance (Peperkamp et al., 2010; Heisterueber et al., 2014; Lialiou et al., 2023; also discussed in Dupoux et al., 1997), which might be a reflection of general recall abilities, such as overall memory or auditory processing abilities. Those might naturally vary between participants and could also contribute to the baseline differences in phoneme recall mentioned above.

1.4. Research question and hypotheses

In spite of the limitations on the strict comparability, it becomes clear from the literature that PSRTs provide a thoroughly developed and cross-linguistically tested experimental approach to investigate the extent to which listeners can memorize word stress patterns. Given that lexical stress storage remains to be investigated for Papuan Malay (Section 1.1), the aim of the current study is to investigate stress pattern recall in listeners of this language. For comparability with the most recent investigations covering typologically different languages, the current study makes uses of the stimulus materials provided in Rahmani et al. (2015) and a new version of the experimental setup developed for Bruggeman (2020) and Lialiou et al. (2023). The details of this approach are further outlined in Section 2. For comparability with well-studied languages, the PSRT in the current study is also carried out with German listeners. German is reported as more representative of Germanic stress than English, mainly due to the acoustic realization and the cues listeners attend to (e.g., Cutler, 2005; Cooper et al., 2002). Duration was reported as the main acoustic correlate of word stress in German (Dogil, 1999). The research question answered in the current study is therefore:

RQ: To what extent do Papuan Malay and German listeners recall stress patterns from memory?

Answering this research question sheds new light on the nature of Papuan Malay stress patterns. These were shown to be readily available in the speech signal and useful to listeners for word disambiguation (Kaland, 2019, 2021; Kaland et al., 2021). However, the distribution of word stress patterns is highly regular and predictable, which demands little to no lexical storage of stress information. It is therefore unclear how well Papuan Malay listeners are able to recall stress patterns, and the hypotheses for this language are therefore twofold. If listeners are highly proficient in recalling stress patterns, it is likely due to their native ability to store stress information. However, if listeners show difficulties recalling stress patterns, this is taken as an indication that their native language does not require them to do so. A rough estimation of the possible range within which the hypothesized Papuan Malay index may fall can be obtained by considering the other languages tested in previous studies. Papuan Malay listeners’ performance in a PSRT could be similar to Polish listeners’ (e.g., Peperkamp & Dupoux, 2002; Peperkamp et al., 2010), as both languages have highly regular word stress patterns and a clearly defined set of exceptional patterns. Note that exceptional patterns differ in nature between Papuan Malay and Polish. That is, in Papuan Malay (exceptional) ultimate stress depends entirely on the segmental makeup of the default stress location and is found in native roots (e.g., Kluge, 2017; Kaland et al., 2021). In Polish, however, (exceptional) ante-penultimate stress occurs mainly in borrowings (e.g., Comrie, 1976). Thus, Papuan Malay listeners most likely rely on an ability to distinguish the two main stress patterns, which have a more native status (i.e., historically longer part of their perception) than in Polish. This hypothesis is furthermore supported by Polish listeners’ tendency to accept penultimate stress in borrowings (e.g., Domahs et al., 2012). It could therefore be that Papuan Malay listeners are slightly better at stress pattern recall than Polish ones. Whether this subtle difference is detectable in a PSRT remains to be seen. Note that the Polish index values varied between 11.6 (Peperkamp & Dupoux, 2002) and 22.9 (Peperkamp et al., 2010), possibly due to task differences such as stimulus material and sequence lengths. It is furthermore expected that Papuan Malay listeners outperform the Indonesian ones in Rahmani et al. (2015), given that the Indonesian standard variety tested in that study is reported to be stressless (e.g., Van Zanten & Van Heuven, 1998). Thus, insofar as the previous studies are comparable, the hypothesized range of the index value for Papuan Malay listeners lies approximately between 11.6/22.9 (Polish) and 35.5 (Indonesian).

By comparison, German listeners are expected to have a better recall of stress patterns than the Papuan Malay ones, as German stress patterns are less predictable and require lexical storage to a larger extent (Domahs et al., 2014). It is expected that German listeners’ performance is similar to the Dutch ones in Rahmani et al. (2015) (see Table 1).

2.1. Participants

In total 26 native speakers of Papuan Malay completed the experiment: 20 female, 6 male; mean age: 24; age range: 18–35. Some participants spoke other languages apart from Papuan Malay: Indonesian (2), Yali (2), Biak (1) and English (1). All of them learned Papuan Malay from birth and reported to speak it in daily life (with family, at school, at work). Participants who did not pass the training round (Section 2.4) or did not provide responses to all test stimuli were not included in the above number and not included for data analysis (8).

In total 26 native speakers of German completed the experiment: 14 female, 12 male; mean age: 31; age range: 20–64. Some of them reported to speak a dialect in addition to their standard variety of German: Bergisch (1), Badisch (1), Fränkisch (1) and Sächsisch (1).

All participants carried out the experiment in a quiet room using headphones. None of them reported having hearing problems.

2.2. Stimuli

Stimuli consisted of 48 audio files with spoken versions of nonce words, balanced for contrast (phoneme, stress), speaker gender (male, female), and speaker language (Dutch, Persian). Each contrast was generated by two nonce words (A and B). That is, the phonemic contrast occurred between ['mu.ku]^A and ['mu.nu]^B, and the stress contrast occurred between ['nu.mi]^A/[nu.'mi]^B. Thus, the phonemic contrast was based on a consonantal difference in the second unstressed syllable: [ku] or [nu]. The stress contrast was based on a combination of cues: duration, intensity, f0, and formant frequencies (with higher values obtained for the stressed syllable than for the unstressed syllable). Note that the choice for Dutch and Persian speakers in the stimulus materials was primarily made in Rahmani et al. (2015), in particular to test the recall of Persian stress patterns using Dutch as a control language. The paradigm was suitable to test other languages as well and, crucially, the stimuli are nonce words in Papuan Malay (not occurring in any of the word lists or conversation transcripts in Kluge 2017), and in German. Detailed acoustic characteristics of the stimuli are described in Rahmani et al. (2015, Table 2) and f0 contours are analysed in detail in Bruggeman (2020, p. 161). There were three tokens of each item, resulting in a total of 4 (items) ×2 (languages) ×2 (genders) ×3 (tokens) = 48 words.

2.3. Design

The PSRT tested the recall of sequences consisting of three, four or five nonce words (five sequences per length). The sequences were designed so that there was either a phoneme or a stress contrast present, and were generated so that repeated items (AA or BB) were different versions of that item, as produced by the speakers. The sequences used for each contrast in the PSRT are listed in Table 2 and are identical to the ones in Rahmani et al. (2015).

2.4. Procedure

Before the start of the experiment, participants were asked which language(s) they spoke to assess their level of native language proficiency and multilingualism. Thereafter, they received instructions about the experimental task. The experiment was run using PsyToolkit (Stoet, 2010, 2017). The procedure described in Rahmani et al. (2015, p. 6) was implemented in the PsyToolkit online environment (HTML-based), such that the experiment could be run remotely via a web browser (Kaland, 2023). Instructions were translated from English into Papuan Malay. The contrasts were tested in two block orders: half of the participants started with the phoneme contrast, the other half with the stress contrast.

Each part consisted of a training session, a warm-up round and the actual test block. In the training session, participants learned to associate member A of the contrast with key “1” and member B with key “2”. This was done by instructing them to press any of the keys and listen to all items associated with that key (12 items for each key). Then, participants pressed any of the keys and listened to a single item chosen randomly from the set associated to that key. They could press either key as often as they wanted until they indicated that they had learned the item-key associations. After that, participants heard a single item, which was randomly chosen from the set, and pressed the key that they just learned was associated to that item. Feedback was given on whether participants correctly or incorrectly identified the item in order to facilitate their learning of the button-response association. After they had identified eight items correctly, the training session ended.

In the warm-up round participants listened to sequences of two nonce words (AA, AB, BA, BB, for each language set). Their task was to identify the sequence by pressing the correct key combination. They needed to complete all eight sequences without errors before they could continue. Each sequence was presented (repeatedly) until participants had correctly identified it.

In the test block, participants identified sequences of three, four or five nonce words (Table 2), which were presented in random order, different for each participant. Within each sequence, items were chosen randomly from the same speaker gender and language and therefore only varied in contrast and version (no items were repeated within sequences). Participants received no feedback on their performance during the test block. A test block consisted of 30 trials (5 sequences × 3 lengths × 2 languages).

2.5. Statistical analysis

A generalized linear mixed model (GLMM) analysis was run on the correctness scores (0 for incorrect, 1 for correct), with contrast (two levels: phoneme, stress), sequence length (successive differences contrast coded, three levels: 3, 4, 5), speaker language (two levels: Dutch, Persian), speaker gender (two levels: male, female), block order (two levels: phoneme first, stress first) experiment language (two levels: Papuan Malay, German) as fixed factors, and with participant and item as random intercepts. The following formula was used: glmer(score ∼ contrast + seq.length + spk.lg + spk.gender + block.order + exp.lg + (1|item) + (1|pp), data, contrasts = list(seq.length = contr.sdif), family = binomial).

In addition, the index proposed by Peperkamp & Dupoux (2002) was calculated for each sequence length by subtracting the error rate for the phoneme contrasts from the error rate for the stress contrasts.