Bengali nasal vowels: lexical representation and listener perception

1.1 Background

The phonetics and phonology of nasal vowels have been studied extensively and, while an accurate objective measurement of nasality is difficult to obtain, several theoretical approaches have been put forward as to the phonetic characteristics of nasality, the cues in the speech signal marking nasality used by listeners, and their lexical representations (see Krämer 2017: 408ff for a review). Theories of how listeners abstract information from the speech signal and recognise speech sounds and words can be said to differ in two fundamental ways. First, they differ in which information is stored for any given word in the mental lexicon. The theoretical proposals range from models that assume listeners store very rich episodic information for every word every time it is encountered (Johnson 1997; Pierrehumbert 2002) to models that are more parsimonious and store only information that is contrastive and where each word or morpheme only has one underlying representation (e.g. Norris et al. 2000; Lahiri and Reetz 2002, 2010). Second, especially in the latter group, theories differ in terms of which information is extracted from the speech signal to access the mental representations, for example articulatory events (e.g. Liberman and Mattingly 1985; Fowler 1986), phonemes (e.g. Marslen-Wilson 1987, 1989; Norris 1994; Norris and McQueen 2008), or features (e.g. Gaskell and Marslen-Wilson 1997; Lahiri and Reetz 2002, 2010). One of the focal debates amongst the latter models, and one of the questions investigated in this paper, centres on the issue of the specification of nasality in the lexicon and whether the feature [nasal] is privative or equipollent.

Chomsky and Halle (1968: 316) treat [nasal] as a binary feature, proposing nasal [+nasal] and nonnasal [−nasal]. More complex representations have been proposed in recent years where only marked, prominent or active features are represented (e.g. Clements 2001), suggesting that languages differ in their representation. For instance, Cohn (1993) suggested that, since French contrasts nasal and oral vowels, these ought to be represented as [+nasal] and [−nasal] respectively, while this is irrelevant for English vowels. Clements, however, would distinguish between lexical representation and phonologically active features, such that a lexical representation may lack a specification to begin with, but a feature may become active later in the phonology. Similarly, Trigo (1993: 393) proposes, in an investigation of Lusitanian Portuguese, that the feature nasal is equipollent ([+nasal] vs. [−nasal]) in languages where it ‘crucially distinguishes among two or more elements’ (e.g. in Bengali) while it is privative ([nasal] vs. no value) in languages where this is not the case (e.g. in English).^[1] In addition, in equipollent systems, there have also been conflicting proposals regarding the representation of contextually nasalised vowels, with some suggesting that those vowels are represented as [+nasal] (cf. Ohala and Ohala 1995 for English CVN words) or [−nasal] (cf. Chomsky and Halle 1968).

From a theoretical standpoint, an equipollent or binary system which refers to nasal sounds as [+nasal] and consequently assumes that oral vowels and consonants are specified as [−nasal] would presuppose that all items which are specified for either [+nasal] or [−nasal] form a coherent class with certain properties common to them all. It would also follow that both features could be active in phonological processes such as spreading. The assumptions for a privative or monovalent system are rather different: for instance, the opposite of the feature [nasal], inherent to all oral vowels and consonants, does not form a coherent set since it is not specified (sometimes indicated by empty brackets [ ]; Lahiri 2018). There is no one feature which groups together all oral vowels and oral consonants and thus there is no featural information which is extracted from the signal to match with a mental representation of oral sounds or to be active in phonological processes. One example of a theoretical lexical access account which proposes a monovalent system and operates on a feature level rather than using phonemes (unlike, for example the Cohort model; cf. Marslen-Wilson 1987, 1989) is the Featurally Underspecified Lexicon (FUL; Lahiri and Reetz 2002, 2010).

In addition to the question regarding the representation of nasality, this paper also addresses whether listeners distinguish between underlying and contextual nasality, and if so, which cues they use to discriminate between the two. The present study is not primarily concerned with the question of the extent to which the contextually nasalised vowel in a CVN word (e.g. English ban) differs phonetically from an underlying nasal vowel, but with the question of what native listeners of a language with both underlyingly nasal and contextually nasalised vowels do when they perceive any nasality in a vowel. If one assumes that there is a phonetic difference between contextual and underlying nasals (e.g. Cohn 1993; Beddor et al. 2013), hypotheses concerning the interpretation of nasality from the signal would differ for the two types of vowels. If phonetic cues are unambiguous as to the source of the nasality, then listeners would surely make use of these cues and thus would be able to accurately identify the source of the nasality in the signal (i.e., an underlying nasal vs. a contextually nasalised segment). If, on the other hand, representation governs perception and recognition, any nasality in the signal would in the first instance lead listeners to an underlying, fully represented nasal vowel.

There has been strong evidence that listeners are affected by top-down effects when processing the acoustic signal. Both likelihood and lexical status of a perceived signal (e.g. Ganong 1980) as well as the phonological system of a listener’s native language(s) have been found to guide perception and can lead to listeners seemingly disregarding certain acoustic cues (Jongman et al. 1992). Ganong’s seminal paper (1980) showed that categorisation of stimuli with ambiguous VOT could depend on the existence of a word. Thus, the same VOT duration would be categorised as [t] if the decision was between tack/*dack but as [d] if the choice was between *tesk/desk. Jongman et al.’s experiment also revealed that identical vowel duration led to significant variation in choosing between words and nonwords based on their phonological representation. They appended different word-initial consonants to a continuum from [at] to [aːt] to create two pairs of real words in Dutch: /zat/ ∼ /zaːd/ ‘drunk’ ∼ ‘seed’ and /stad/ ∼ /staːt/ ‘city’ ∼ ‘state’. Since Dutch voiced stops are devoiced word finally, the surface pairs would be [zat] ∼ [zaːt] and [stat] ∼ [staːt]. The results showed that the vowel categorisation between long and short depended on the underlying phonological representation of the voicing of the final consonants:

The phoneme boundary for /zat/ ∼ /zaːd/ occurred at a significantly shorter vowel length than /stad/ ∼ /staːt/. Thus, the results showed that the categorisation of ambiguous vowel length was affected by the underlying voicing of the following word final consonant.

The fact that phonological representations guide perception has been demonstrated in studies across different languages as well as dialects (i.e. Pallier et al. 1997; Pallier et al. 2001; Scharinger and Lahiri 2010), which show that phonetically identical contrasts are interpreted differently by speakers of different languages/dialects due to differences in the phonological features that are represented in their lexicon. Scharinger and Lahiri (2010), for example, show how the different specification of short front vowels for [low] in two dialects of English (American and New Zealand) affected listeners’ processing in a semantic priming study.

In the present study we are not, however, dealing with variability across dialects or languages but with the inherently variable phonetic property of nasality. Thus, our central question is whether listeners ascribe surface nasality to regressive assimilation from a following nasal consonant or whether an interpretation as an underlying nasal vowel would be equally, or indeed more, likely in languages such as Bengali, which have both contextually nasalised and underlyingly nasal vowels. Thus, in addition to the question of how nasality is represented, we are also concerned with the question whether listeners’ choices are indeed governed by this representational information. While there has been some previous work in this area (cf. Section 1.3 below), we are using different methods to elicit quicker decisions from listeners without any context in order to shed light on automatic access to stored representations.

1.2 Acoustic properties of vowel nasality

1.3 Lexical representation of nasality

Lexical access models predominantly focus on access routes to the stored information and often do not provide much detail about the representation of features, such as nasality, or phonemes but only present a brief overview of the phonological information contained in the mental representations (see McQueen 2005: 268 for an overview of proposed representations in common speech perception models). Models such as TRACE (Elman and McClelland 1986; McClelland and Elman 1986) or Cohort (Marslen-Wilson and Welsh 1978; Marslen-Wilson 1993) and the Distributed Cohort Model (DCM, Gaskell and Marlsen-Wilson 1997) as well as Stevens’ (2002) Acoustic Landmarks Model assume a featural representation either at the pre-lexical stage or as underlying representations. An account which assumes featural underspecification would be compatible with models proposing featural representations, while models such as Shortlist A and B (Norris 1994; Norris and McQueen 2008) and exemplar-based models (Johnson 1997; Pierrehumbert 2002), which propose phonemic representations, would result in different processing predictions. The present study is primarily concerned with how contrastive features are stored and we provide predictions based on underspecified representations, phonemic representations, and exemplar-based representations in detail below.

As far as the question of the representation of vowel nasality is concerned, earlier research on Bengali shows support for [nasal] as a privative feature, claiming that predictable nasality is not represented (Lahiri and Marslen-Wilson 1991, 1992; henceforth L&M-W). Contextually nasalised vowels in CVN sequences would thus be represented without nasality, e.g. [bɑ̃n] ‘flood’ from underlying /bɑn/. This results in ambiguous surface forms in languages like Bengali where there are also underlying nasal vowels. These cases can only be disambiguated by the following consonant. In the case of an oral vowel in the sensory input (e.g. [bɑd̪] ‘omit’), an underlyingly nasal vowel would no longer be a candidate since this combination results in a no-mismatch between the sensory input (underspecified) and the representation ([nasal]). Nasal or contextually nasalised vowels in cases like Bengali (e.g. [bɑ̃d̪ʰ] ‘dam’, [bɑ̃n] ‘flood’), however, should not mismatch with the underlyingly oral vowel since oral vowels have no featural specification for nasality.

L&M-W (1991) used a gating task (‘t Hart and Cohen 1964; Grosjean 1980) with incremental presentation of the CV segment of CVC words where listeners were asked to respond with a full word which forms a continuation of the phonetic CV sequence they were perceiving. Their stimuli were sets of triplets in Bengali (CVC [bɑd], CVN [bɑ̃n], and CṼC [bɑ̃d̪ʰ]) as well as doublets where no word with a nasal vowel exists in the same consonantal environment (CVC [lobʰ] ‘greed’ and CVN [lõm] ‘body hair’).

L&M-W found that in the triplet condition both stimuli with contextualised nasal vowels and underlying nasal vowels, i.e. CVN and CṼC stimuli, respectively elicited a large percentage of CṼC responses at the offset of the vowel.^[7] This shows that listeners take nasalisation as a cue that they have heard a CṼC rather than a CVN word. CVN responses for both categories were very low (below 8%) but this changes rapidly as soon as consonantal information becomes available since this disambiguates the surface nasality. The best match for the nasality in the auditory signal was a word with a vowel underlyingly specified for nasality, i.e. a CṼC word. CVN words are still an option since there is no mismatch between the signal and the underspecified underlying representation of the oral vowel. This data shows support for underspecification since a surface hypothesis would have predicted an equal distribution of CṼC and CVN responses here as both vowels would have been specified for nasality.

CVC stimuli resulted in a large proportion of CVC responses (>80%) even at very early gates.^[8] CṼC responses were almost non-existent (0.7%) but there are a relatively large number of CVN responses (13.4%) considering the lack of nasality in the signal (L&M-W 1991). This phenomenon is difficult to explain with a surface representation account since, with nasality present in both representations, there should be no difference between CVN and CṼC responses to CVC stimuli. This pattern is more easily explained by proposing an underlyingly oral vowel in CVN words. CVN responses to CVC stimuli are thus not only permitted but, based on the underlying representation alone, should be as likely as CVC responses since they have the same representation (L&M-W 1991). The fact that there are more CVC responses than CVN responses corresponds to the proportion of CVC and CVN patterns in the language, with CVC sequences being considerably more frequent than CVNs.

Data from L&M-W’s doublet sets exhibited similar patterns for the CVC stimuli, and the CVN stimuli result in a different pattern where listeners produced 65% CVC responses, 16% CVN responses and 17% CṼC responses even though no CṼC item, in this consonantal frame, is available in the lexicon of the language. In these cases, participants often responded with CṼC words from Hindi, another language they were familiar with. The increase in CVC and CVN responses compared to the triplet set shows that the information in the signal matches both representations and there is no other competitor containing nasality (unlike in the triplet case). This makes the distribution of responses more similar to the CVC stimuli responses above, which are in part determined by pattern frequency in the language.

In a subsequent study using Hindi and English, Ohala and Ohala (1995) attempted to replicate L&M-W’s findings with methodological modifications since they had several criticisms of the original study, including the fact that in L&M-W’s study no statistical analysis was possible since participants’ responses were not constrained, and that previously heard versions at earlier gates could have influenced participants’ responses. Ohala & Ohala restricted the participants’ responses and, in order to prevent previous stimuli affecting subsequent responses, presented only one gated version of each stimulus per participant (with the exception of the most severely gated stimulus, which was followed by the full word after four trials). While they found results similar to those of L&M-W (1991) in their Hindi triplet condition as well as with CVN stimuli in the Hindi doublet condition, thereby lending support to the underspecification hypothesis, they do not see either gating study as providing convincing evidence for underspecification of nasality but state that ‘at best [these findings] are compatible with both the UR [underlying representation] and SR [surface representation] hypotheses’ (Ohala and Ohala 1995: 57).

Both studies (Lahiri and Marslen-Wilson 1991, 1992; Ohala and Ohala 1995) thus showed that in a gating paradigm listeners displayed an overwhelming preference for contrastively nasal vowels in response to any nasality in the signal (CVN or CṼC), to the extent that, in cases where there was a lexical gap (e.g. in Lahiri & Marslen-Wilson’s doublet conditions), listeners provided words with nasal vowels in response to CVN fragments from another language they were familiar with (Hindi).

1.4 Present study and predictions

To address the central questions of this study – whether nasality is represented as an equipollent or a privative feature, and how listeners use phonetic and phonological nasality in processing – we conducted a series of experiments using a cross-modal (audio-visual) form priming task with a forced-choice response paradigm (Cooper et al. 2002). All stimuli (described in greater detail in Section 2.1) were monosyllabic Bengali words belonging to one of three types: CṼC, CVN and CVC. The CV/Ṽ segments of these words were used as primes, with the full words as targets.

Before setting out the predictions resulting from different types of representational hypotheses, we briefly discuss the distributional patterns of nasal and non-nasal vowels in Bengali.

The Bengali vowel system consists of seven oral vowels phonemes and seven nasal counterparts (cf. Chatterji 1926; Ferguson and Chowdhury 1960). All oral vowels can also become nasalised before the three nasal consonants /m, n, ŋ/. Thus, oral vowels are, evidently, the most frequent as they can (theoretically) occur with any of the language’s 30 oral consonants. While nasal vowels can occur preceding all consonants apart from the nasals, they are most frequent before coronal consonants. As contextually nasalised vowels can only occur in certain environments (preceding one of the three nasals in the language), these are the most restricted and therefore least likely. However, these frequency assessments, based on the combinatorial possibilities afforded by the phonological inventory of the language, only tell part of the story, since it also matters how frequent the words are in which these patterns occur. Many of the words with the CVN pattern are highly frequent even though the combinatorial possibilities are the most severely restricted.

Depending on the representational specifications (i.e. privative vs. equipollent) as well as the precise specification of vowels in CVN contexts and the surface cues in the acoustic signal, several hypotheses are possible. We outline below one set of surface (phonetic) predictions (cf. Figure 1) based on the acoustic cues (assuming no difference between contextually nasalised and underlyingly nasal vowels) and one based on a monovalent system where CVN is unspecified for nasality (cf. Figure 2). In a surface account, where listeners are able to directly match primes and targets based on purely phonetic cues, we would expect faster latencies only for identity pairs since this approach would result in only one possible match per prime (cf. Figure 1).

Figure 1:

Surface phonetic predictions for CṼ(C), CV(N) and CV(C) primes.

Figure 2:

Phonological predictions for match and no-mismatch relationships for CṼ(C), CV(N) and CV(C) primes in an underspecification account.

In a monovalent system, however, where nasal vowels (CṼC) are specified as [nasal] while oral vowels (CVC) and vowels nasalised due to a following nasal (CVN) are not specified for nasality, shorter response latencies as well as lower error rates should be observed after CṼC primes (compared to CVC and CVN primes) since their representation better matches the surface form. In Figure 2, we lay out in greater detail the predictions made by an underspecification account assuming a monovalent feature [nasal]. As far as this approach is concerned, the only genuine match is that between a surface nasal vowel (from either CṼC or CVN) and an underlyingly specified nasal vowel /CṼC/ (cf. Lahiri and Reetz, 2002, 2010). All other conditions should result in a no-mismatch scenario between the input and the underlying representation since no nasal or oral feature is specified in either CVC or CVN words.

This results in predictions of only two real match conditions across the possible prime-target combinations. We expect an identity match between the underlying nasal vowel prime (CṼC) and its identity target and another match between a surface nasal vowel from a CVN prime and the specified nasal vowel target (CṼC). Faster response latencies and lower error rates are predicted for match conditions than for no-mismatch conditions (Lahiri and Reetz 2002). In terms of the error data, however, it is unclear whether the distinction between match and no-mismatch will surface clearly. Surface nasalisation may well play a more crucial role in the error data than in the reaction times, which are a more gradient measure. Since nasality is highly salient, this may lead to more accurate matching in cases where the prime contains a nasal or contextually nasalised vowel and a word with either option (CṼC or CVN) is present as a target.

In approaches which are fundamentally guided by frequency and experience (such as an exemplar model), it is never precisely articulated how one particular exemplar is chosen out of a number of possible options. Frequency will contribute to this selection as will, presumably, surface cues, such as the degree of nasalisation of a vowel. Since the degree of nasalisation can vary greatly in both underlying nasals and contextual nasals and exact matches are not required in an exemplar model, there must be a gradient approach and nasalisation in the signal could lead to either an underlying nasal or a contextually nasalised vowel, depending on other contributing factors such as, for instance, frequency. It is thus difficult to provide clear-cut predictions for such an exemplar model: these predictions would fall somewhere in between those made by a fine-grained phonetic account and an underspecification approach, but they would be clearly distinct from both. Phonetic accounts would only lead to an identity match, and an underspecification account makes the clear prediction that, given any nasality in the signal, the first choice would be an underlying nasal. as this is the only option specified for nasality.

In terms of frequency, two different aspects need to be considered. Firstly, if nasal vowels are heard more frequently in a CṼC rather than a CVN frame (as there are only three nasal consonants in Bengali which would trigger regressive nasalisation), listeners should be faster to match nasal vowel primes (be it from a CṼC or CVN prime) to CṼC targets. However, since the oral vowel only occurs in CVC contexts, matching these to CVC targets should be fastest overall since this is the most frequent pattern. If the number of neighbours is a determining factor here, a faster response might be expected for CṼC prime-target combinations as these have fewer neighbours. CVN words have even fewer neighbours and should thus result in even faster reaction times than CṼC words, with CVC words being reacted to slowest.

There is no published corpus for Bengali to allow us to compute neighbourhood density and vowel frequency statistics. Basic lexical frequency counts were taken from a Bengali corpus of three million words (Dash and Chaudhuri 2001) and mean frequencies for each condition can be found in Table 1. However, these can be misleading as the lexical distribution is not the same. The number of words with nasalised vowels in the context of nasal consonants /m n ŋ/ is, by definition, smaller than that of words with oral vowels which can be followed by any of the 30 oral consonants. However, in contrast to (for instance) French, and other languages where nasality has been studied, Bengali is, to our knowledge, one of the few languages which has a full set of contrastive nasal and oral vowels (seven each) (Ferguson and Chowdhury 1960: 32; Hajek 2013; Klaiman and Lahiri 2018: 423). Consequently, Bengali provides the best possible foundation for an investigation of the phonological representation of nasality.

2.1 Stimuli

2.1.1 Primes

CV fragments of 42 common monosyllabic CVC words of Bengali were used as primes and they were divided into three sets of near-minimal pairs: one set of 14 triplets and two sets of 14 doublets each (for examples see Table 1). The triplet set contained CVC strings in which all possible variations of vocalic nasality led to real words in the language: oral vowels (e.g. [tʃɑl] চাল ‘unboiled rice’), nasal vowels (e.g. [tʃɑ̃d̪] চাঁদ ‘moon’) and nasalised oral vowels due to a following nasal (e.g. [tʃɑ̃n] চান ‘bath’). In the first doublet set (NoCṼC), words with the same consonant sequence containing a nasal vowel do not exist in Bengali and thus the two primes in this set contained oral vowels (e.g. [t̪il] ‘sesame seeds’) and nasalised oral vowels due to a following nasal (e.g. [t̪̪̪ĩn] ‘three’). The second doublet set (NoCVN) contained only oral vowels (e.g. [dʒ^hɑl] ঝাল ‘spicy hot’) and underlying nasal vowels (e.g. [dʒ^hɑ̃p] ঝাঁপ ‘jump’), and in these cases the corresponding minimally different CV pairs with nasalised oral vowels are not attested in Bengali.

All primes were recorded in full and then truncated using PRAAT (Boersma and Weenink 2011) to include the complete duration of the vowel while ensuring no consonantal information is available. This was done based on the spectral properties; obstruents were cut at the beginning of the closure and for sonorants, decisions were based on close examination of the spectrogram as well as our own perception (cf. Figure 3 for an example of full words and truncated primes). There were no significant length differences of the individual vowels between the three conditions (oral, nasalised, underlyingly nasal).

Figure 3:

Example of full CṼC word /ʃõk/ (left panel) and CVN word /ʃon/ (right panel). The lines mark the ends of the prime fragments /ʃõ/ and /ʃo/ respectively.

2.1.4 Acoustic measurements of nasality

As discussed in Section 1.2.1, we believe that Q1 is the most reliable vowel- and speaker-independent acoustic correlate of nasalisation. However, since other measures have been preferred in the literature, we therefore report the A1–P0 compensated and A1–P1 compensated measures provided by Styler (2017) as well as the Q1 measure. We computed these values for each vowel at five points throughout the vowel (at 20, 40, 60, 80 and 100% of the vowel duration) using a modified PRAAT script (Boersma and Weenink 2011) from Styler (2017).^[9]

The measure A1–P0 compensated grouped the nasal and oral vowels together as oral, while the nasalised vowels were calculated as being the ‘most nasal’, and A1–P1 compensated could not differentiate between the three different vowel types at all. Table 2 shows the results of these three measures at five positions in the vowel for all three different stimulus types (CVN, CṼC and CVC) as well as the results of t-tests to compare all possible combinations within a measure. For the same stimuli, the Q1 measure groups the nasalised and the underlying nasal vowels together, differentiating them from the oral vowels (Figure 4).

Table 2:

Average Q1, A1P0 compensated, and A1P1 compensated measures at five timing points for all vowels in CṼC, CVN, and CVC stimuli and results of t-tests (Tukey HSD) comparing those measures (* indicates significance).

		20%	40%	60%	80%	100%
CṼC	Q1	3.42	2.86	2.59	2.71	3.07
CVN	Q1	3.02	2.51	1.76	1.90	2.66
CVC	Q1	5.98	4.30	4.13	4.47	4.52
CVN/CṼC	t	p = 0.9031	p = 0.8803	p = 0.3798	p = 0.3276	p = 0.8207
CVN/CVC	t	p = 0.0018*	p = 0.0251*	p = 0.0002*	p < 0.0001*	p = 0.0090*
CṼC/CVC	t	p = 0.0079*	p = 0.0090*	p = 0.0021*	p = 0.0029*	p = 0.0521
CṼC	a1p0 comp	0.34	1.24	0.26	0.10	−1.21
CVN	a1p0 comp	−0.15	−0.76	−1.34	−3.51	−2.97
CVC	a1p0 comp	4.13	1.40	0.61	0.72	0.58
CVN/CṼC	t	p = 0.9587	p = 0.1917	p = 0.3098	p = 0.0219*	p = 0.3987
CVN/CVC	t	p = 0.0251*	p = 0.1001	p = 0.1282	p = 0.0022*	p = 0.0140*
CṼC/CVC	t	p = 0.0538	p = 0.9863	p = 0.9355	p = 0.8685	p = 0.3235
CṼC	a1p1 comp	15.85	17.85	18.49	18.82	17.34
CVN	a1p1 comp	15.80	17.78	20.45	19.11	23.63
CVC	a1p1 comp	19.37	17.29	18.60	19.31	21.78
CVN/CṼC	t	p = 0.9999	p = 0.9998	p = 0.8853	p = 0.9972	p = 0.1713
CVN/CVC	t	p = 0.5085	p = 0.9861	p = 0.8779	p = 0.9984	p = 0.8278
CṼC/CVC	t	p = 0.5176	p = 0.9817	p = 0.9995	p = 0.9904	p = 0.3460

Figure 4:

Average Q1, A1P0 compensated, and A1P1 compensated measures at five timing points for all vowels in CṼC (nasal), CVN (nasalised), and CVC (oral) stimuli.

This data show that there are only small differences in the Q1 values between the vowels of CṼC and CVN stimuli at any of the timing points (not even at 20% or 40%, where CVN vowels may be expected to show less nasality) while the Q1 values differ at all timing points between oral (CVC) vowels and both underlyingly nasal (CṼC) and contextualised nasal vowels (CVN). This demonstrates that there are no significant acoustic differences in the degree of nasalisation between CṼC and CVN stimuli, while both are acoustically distinct from CVC items. We further analysed the formant transitions of F1 and F2 as well as how frequently a target with a specific final consonant was selected for a given input fragment which determined that there were no additional consistent phonetic cues which could influence the main result. The results of the formant transitions show a considerable number of differences for F1 (as indicator of manner of articulation) between oral vowels and nasal vowel. For F2, there are only few differences, mostly for high vowels (which are few in the experiment). Thus, the coarticulation points to different manners of articulation, while differences in place of articulation are not evident in this formant analysis. The results of the analysis based on a targets’ final consonant indicate that participants choose alternative words mostly on the basis of manner of articulation rather than based on place information. This contradicts the information available from the formant transitions. Detailed results for these analyses can be found in the Supplementary material to this article.

2.2 Procedure

In the priming experiment, participants were tested in groups of at most 16 in a quiet and darkened room. The auditory primes were played through individual on-ear headphones (SONY MDR110 LP) and visual targets were projected onto a screen in Bengali script with a data projector. Reactions and reaction times were captured using a custom-made multi-participant experimental setup (Reetz and Kleinmann 2003). In a standard forced-choice paradigm, participants were asked to indicate whether the CV fragment they heard belonged to the left or right target word on the screen. No information was given about the oral or nasal quality of the signals. Responses were made via individual custom-made two-button boxes with the left button corresponding to the left target and the right to the right target. All participants were right-handed and used their dominant hand to indicate if the prime matched the target displayed on the right. A ten-item practice task with items not used in the main experiment was conducted and re-run if necessary until it was clear participants were comfortable with the task.

Trials were separated into two blocks, which were presented to different participants. Every prime was presented once in the case of the doublets and three times in the triplet set in each of the two blocks, to ensure that each possible prime-target combination was only shown once per block (cf. Table 1). Each block consisted of 182 pseudo-randomised trials presented in a different order in each block (Table 1 shows all 13 combinations in each block for one set of the 14 items). The side of presentation (left vs. right) of the target words was randomised and counterbalanced across blocks (cf. Table 1). Participants heard one beep followed by a 200 ms silence before each auditory prime. Immediately at the end of the prime the visual targets (pairs of words) were displayed for 800 ms. After an additional 700 ms (i.e., 1500 ms after the end of the prime and the beginning of the visual display) the next trial started with its beep. After every 14 trials a sequence of three beeps was presented, which served as a short break. Participants were told they would hear the beginning of a word and then had to decide which of the words presented on the screen was the closest continuation for what they had heard. They were instructed to respond as quickly and accurately as possible.

2.3 Word comprehension tasks

Two word comprehension tasks were conducted after the priming experiment to test listeners’ perception of the stimuli used in the that experiment. In both cases, listeners were presented with the same CV/Ṽ fragments used in the priming experiment, presented in a completely randomised order. The auditory fragments were played through headphones at a fixed pace. Participants were asked to respond in writing (pen and paper task) with the syllable they heard (Test 1; n = 26) or the complete monosyllabic word the fragment was taken from (Test 2; n = 26).^[10] The results of these tasks are presented in Section 3.1 below.

2.4 Analysis of priming experiment

The data from two participants had to be omitted due to a malfunctioning button box and one additional participant was removed from the analysis due to consistently fast and even negative reaction times suggesting they were reacting to the prime rather than the target. In addition, all reaction times ≤0 ms and outside ±2 standard deviations were excluded from the following analyses (4.24% of data for triplet conditions and 3.96% for doublets). Errors (response accuracy) were analysed first separately, comparing the proportion of correct responses to a chance proportion of 50% using Pearson’s Chi Squared tests. Errors were then analysed using logit generalised linear models, treating responses (error/correct) as a binomial distribution. Reaction times were analysed using linear mixed effects models.

3.1 Word recognition task results

As can be seen in Figure 5, CV syllable and CVC words result in the most accurate responses (91.4 and 75.3%), while CV(N) syllables seem to be most difficult for listeners to identify accurately from CV(N) fragments (2.6%). Even CVN words gave only 34.6% correct responses, which may well be modulated by the words available in the lexicon and their frequencies. This indicates that, in the syllable task, the nasality which is present in the signal of a CV(N) fragment is not unambiguous and is interpreted by listeners as either a CV or a CṼ syllable (49.4 and 48% responses).

Figure 5:

Predictions and error results for triplet match (% correct) and neither (% chosen) conditions.

For CV and CṼ syllables and words, listeners are more likely to choose the correct word rather than any other and while there are a significant number of CṼ syllable and CṼC word responses (36.7 and 30.8%) to CṼ fragments, there are hardly any or only few CV(N) responses (0.3 and 16.2% respectively).

3.2 Priming triplet results

Since, as mentioned above, in one third of the trials there was no possible identity match of prime and target, the triplet data will be analysed in two separate conditions: match and neither. The match condition includes all instances where an exact match to the prime was available as a target (e.g. prime [tʃɑ̃] (from [tʃɑ̃d̪]); targets: [tʃɑ̃d̪] – [tʃɑl]) and here only correct trials are included in our analysis. The neither condition contains the data from those trials where no identity match was available (e.g. prime [tʃɑ̃] (from [tʃɑ̃d̪]); targets: [tʃɑ̃n] – [tʃɑl]) and therefore, since there is no correct item, all trials are analysed.

In the match condition, errors were analysed using a logit generalised linear model treating responses (error/correct) as a binomial distribution, for the fixed effect of prime vowel (CV(N), CṼ, CV), with random intercepts specified for participants and targets. Reaction times were analysed using a linear mixed model for the fixed effects of prime vowel and context of response (nested under prime vowel, i.e., the target that did not match the prime; Prime ∼ Context: CV ∼ CṼC, CV ∼ CVN, CṼ ∼ CVC, CṼ ∼ CVN, CV(N) ∼ CVC, CV(N) ∼ CṼC), with random intercepts specified for participants and targets.

In the neither condition, response preferences were analysed using a logit generalised linear model treating responses (possible response A/possible response B) as a binomial distribution, for the fixed effect of prime vowel, with random intercepts specified for participants and targets. Reaction times were analysed using a linear mixed model for the fixed effects of prime vowel and context of response, with random intercepts specified for participants and targets.

3.2.2 Error analysis

Overall, in the match condition, participants responded with the identity target 61.4% of the time and performed significantly better than chance in this task (χ ²(1) = 208.21, p < 0.001). In analyses by prime vowel, participants’ responses were significantly more accurate than chance in all three conditions: CṼC (χ ²(1) = 189.89, p < 0.001), CVN (χ ²(1) = 26.52, p < 0.001) and CVC (χ ²(1) = 36.70, p < 0.001).

In a logit generalised linear model (fixed effect: prime vowel), the data shows a significant effect for prime vowel (χ ²(2) = 48.06, p < 0.001). In a planned comparison of the percentages of correct responses to targets we find a significant difference between responses following CṼ versus CV(N) primes (χ ²(1) = 40.16, p < 0.001) as well as between those after CṼ versus CV primes (χ ²(1) = 31.90, p < 0.001) with a significantly larger percentage of correct responses to targets following CṼ primes. There is no difference between correct responses after CV(N) and CV primes (χ ²(1) = 0.45, p = 0.500).

Since there is no correct or incorrect response in the neither condition, it is not possible to provide an overall analysis of errors to see whether participants performed above chance. The results (illustrated in Figure 6; Std Errors can be found in Appendix B) were analysed by prime vowel and participants showed a significant preference for one target over the other in all three conditions. When presented with a CṼ (nasal vowel) prime, participants responded with a significantly greater number of CVN targets (63.8%) than CVC targets (36.2%; χ ²(1) = 51.16, p < 0.001). In the CV(N) (nasalised vowel) prime condition, CṼC targets (66.2%) were significantly more frequent than CVC targets (33.8%; χ ²(1) = 70.51, p < 0.001) and when presented with a CV (oral vowel) prime, participants responded with CVN targets 57.4% of the time over CṼC targets, which were chosen significantly less frequently (42.6%; χ ²(1) = 14.56, p < 0.001).

Figure 6:

Predictions and reaction time results for the triplet condition

In a logit generalised linear model, we again see a significant effect for prime vowel (χ ²(2) = 11.91, p < 0.003). The planned comparisons show that participants display a greater bias towards the specific target vowels above after both CV(N) and CṼ primes than after CV primes (CV(N) versus CV: χ ²(1) = 11.10, p < 0.001; CṼ versus CV: χ ²(1) = 5.89, p = 0.015). There is no difference in the degree of preference of the target vowels after CV(N) and CṼ primes (χ ²(1) = 0.81, p = 0.368).

3.3 Priming doublet results

For the doublet analysis, the analyses were split into two separate models – NoCVN, where no word with a nasalised vowel exists in Bengali, and NoCṼC, where no word with a nasal vowel exists in Bengali.

In the NoCVN condition, errors were analysed using a logit generalised linear model treating responses (error/correct) as a binomial distribution, for the fixed effect of prime vowel, with random intercepts specified for participants and targets. Reaction times were analysed using a linear mixed model for the fixed effect of prime vowel, with random intercepts specified for participants and targets.

3.3.2 Error analysis

The overall errors in the two doublet sets were 34%. The data in the doublet condition shows clearly that while listeners detect the identity match well above chance in all conditions (χ ²(1) = 239.10, p < 0.001) and they performed equally well in both doublet sets (χ ²(1) = 0.01, p = 0.922) there are significant differences between the different prime vowels in both reaction times and errors (cf. Figure 7; Std Errors can be found in Appendix B).

Figure 7:

Reaction time results (ms) and error rates for doublet conditions.

In the analysis of prime vowels in the set in which no word with a nasalised vowel exists in Bengali (NoCVN), the data shows a significant difference between the number of errors made following CṼ primes versus CV primes (χ ²(1) = 30.82, p < 0.001). Reactions to targets preceded by CṼ primes were substantially more accurate (26.3% errors) than those observed when primes contained an oral vowel (CV; 41.5% errors).

In the NoCṼC set, where a corresponding target with a nasal vowel did not exist in Bengali, participants made significantly fewer errors (χ ²(1) = 4.56, p = 0.033) when reacting to primes with nasalised vowels before a nasal consonant (CV(N); 31.1%) than when the targets followed oral vowel primes (CV; 37.1%). Furthermore, despite the significantly higher error rates for targets following CV primes, participants still performed significantly above chance (χ ²(1) = 38.16, p < 0.001).