Nineteen native speakers of Spanish (4 males, 15 females, M _Age = 24.3, SD = 8.8), recruited among students of the Complutense University of Madrid, took part in the study. They did not receive any reward for their participation, and none of them was aware of the purpose of the investigation. Participants were monolingual speakers with no relevant differences in their Castilian Spanish dialect, except for 2 subjects, who were bilingual speakers of Catalan, although their Spanish was not different from the standard Castilian variety (see Appendix A1 for details).

The speech material was made up of paired sentences, each with a different lexical affiliation of the target consonant. Although the two-word sequences where the target consonant occurred were identical, the sentences themselves did not form minimal pairs, e.g. Probablemente marcos aluda a Paco para despistar (‘Probably Marcos alluded to Paco to mislead you’) vs. Finalmente marco saluda a Lorena desde el tren (‘Finally Marco greets Lorena from the train’). Furthermore, the prosodic structure was similar between both members of each pair of sentences regarding the strength of the prosodic boundary between the two words forming the minimal pair. The stimuli included 8 pairs of sentences for /s/, 7 for /n/, and 4 for /l/ (see Appendix A2 for details). For every participant 2 repetitions of each uttered sentence were chosen, i.e., those sentences containing a disambiguating pause were rejected, which made up a total of 600 uttered sentences for the target consonant /s/, 524 for /n/, and 300 for /l/, which added up to 1424 utterances.

Speakers were recorded uttering each member of the paired sentences separately in one of two recording sessions, which took place 3 weeks apart from each other so that they did not read both sentences belonging to a pair in the same session. In each session participants were presented with a total of 55 sentences to read out, which included 35 filler sentences next to the 20 stimuli containing the 3 target consonants. Before the recording, participants were allowed to get familiar with the speech material, and it was made sure that they understood all the words in the sentences presented to them. In addition, a trial in which they practiced with some of the sentences was used to provide them with instructions in order to prevent them from producing unnecessary pauses between words, i.e. they were asked to read out each sentence in a continuous manner, so that their speech production was connected. If they made a mistake while recording, they were encouraged to correct themselves by repeating the entire sentence. Besides, 2 repetitions of each sentence were recorded in the same session. The acoustic data were collected at a sampling rate of 44.1 kHz in Praat (Boersma and Weenink, 2018) by means of an AKG C3000 condenser microphone placed at about 30 cm in front of the speakers and connected to a Roland Rubix24 Audio Interface.

The segmentation of the acoustic data was manually conducted in Praat for the two repetitions of each sentence or for the only repetition of the sentence that was available in certain cases. The target consonants were segmented together with their flanking vowels according to the following criteria. The onset of /s/ was defined as the onset of frication noise appearing in the region of 3-5 kHz and higher, as done by Strycharczuk and Kohlberger (2016); while the offset was made to coincide with the appearance of the second formant belonging to the following vowel. For the target consonants /n/ and /l/, which occurred intervocalically, it was both the visible change in intensity and the change in the formant structure respect to the preceding and following vowels which determined the criteria for the segmentation. The onset of V1 was established according to the following criteria: the beginning of F2 when the vowel occurred after a stop or a fricative; the change in intensity after a tap, a trill, or a nasal; both the change in intensity and in the formant structure after a spirant; and finally, after another vowel, the boundary was set considering both the change in formant structure together with the first cycle of rising intensity after a visible decrease in the waveform. The boundary between V2 and the following segment was defined along similar criteria depending on its manner of articulation; in the case of stops or fricatives, the boundary was set when silence or noise, respectively, appeared in the region of 3-5 kHz (Figure 1).

By means of a Praat script, the duration of the three target consonants, as well as that of the preceding vowel (V1) and of the following vowel (V2), was measured based on the segmentation detailed above.

In order to statistically analyze our acoustic data, different linear mixed-effects regression models were separately fitted for each consonant using the lme4 (Bates et al., 2015) package on R (R Core Team, 2021). The acoustic measurements conducted on the data were independently modeled with the categorical variable of condition -i.e. derived onset and canonical onset- as a fixed effect. Random effects were declared for speakers and paired items, since the study was designed with each individual sentence being paired with a prosodically equivalent one in order to control for any confounding variable that might have affected the speakers’ production. This allowed to compare the derived onset condition within each pair of sentences with its canonical onset counterpart, both having the same context and characteristics. In addition, the variance-covariance matrix was fully declared as suggested by Barr et al. (2013), thus including intercepts for both random effects as well as a random slope for condition in by-subject random effects.

In order to assess whether bilingual participants presented any differences with respect to monolingual speakers, we followed Strycharczuk and Kohlberger (2016) and added Bilingual as an additional fixed predictor and its interaction with Condition. Since the effect of Bilingual and its interaction was not significant in any case, we decided to keep these participants in our analyses and did not include these variables in the models reported in the Results section. The dataset analyzed for this study and the R code can be found in https://osf.io/zbems/

The duration of the three target consonants, as well as that of their two flanking vowels, was analyzed on a total of 1424 uttered sentences. Our results replicated previous findings and confirmed that consonant duration is a consistent cue to resyllabification, i.e. in resyllabified contexts, derived onsets are produced with a shorter duration than canonical onsets. This subphonemic difference was not only observed for /s/-as had been firstly reported by Hualde and Prieto (2014) and later by Strycharczuk and Kohlberger (2016)-but also for two other consonants in Castilian Spanish such as /n/ and /l/.

Acoustically, it has been pointed out that derived onset /s/ presents a duration intermediate between canonical codas and canonical onsets (Strycharczuk and Kohlberger, 2016), which suggests that it may constitute an intermediate phonological category. The fact that the acoustic duration of derived onsets was found to be systematically shorter than their canonical counterparts in our results is in line with the theoretical assumption that at least one acoustic cue exists that unifies the behavior of any segment-whatever its identity-for the sheer reason of being resyllabified. In this sense, regardless of the structural interpretation that resyllabified segments may be given, they constitute one phonological object and, as such, they are expected to have some phonetic behavior in common. More precisely, considering the prosodic nature of this phonological object, such a common cue is very likely to be suprasegmental. This finds support in our results, where the duration of the target consonant was affected by the derived condition independently of the consonant identity, especially since consonants originally associated to a coda position in their lexical forms are expected to be realized with a shorter duration than canonical onsets due to the weak nature of codas.

In our results, the difference between derived onsets and canonical onsets was larger for /s/-6.8 ms-than for the other target consonants. In contrast to the results obtained for /s/, the difference between derived onsets and canonical onsets for /n/ was 2.9 ms, while the difference of target consonant /l/-i.e. 5.0 ms-came closer to the results obtained for /s/. These temporal differences, albeit small, might potentially serve as a cue for the segmentation of continuous speech, although this should be further investigated with perceptual experiments.

Such temporal reduction of derived onsets with respect to canonical onsets makes it possible to consider that derived onsets present coda properties. As has been widely shown, gestures in coda position are in antiphase relation, which is inherently more unstable, and this fact renders coda segments shorter than their onset counterparts (Browman and Goldstein, 1992, 1995; Goldstein et al., 2006, p. 229). Another piece of evidence is offered by Byrd et al. (2009), who observed that the articulatory gestures of /n/ are synchronous as canonical onsets, but follow a sequential order when as derived onsets, as happens in coda position.

The evidence presented thus far shows that these consonants retain coda properties as expected from their lexical affiliation, and this might entail that resyllabification has not taken place. Admittedly, we cannot compare our data with canonical codas, since we did not include them in our design. However, we can provide evidence supporting that these consonants present onset properties, which allows to interpret that a post-lexical reorganization of syllable structure has actually occurred. In the case of Castilian Spanish, Strycharczuk and Kohlberger (2016) report a three-way distinction between /Vs#C/, /Vs#V/, and /V#sV/, where derived onsets display intermediate properties with respect to the other two. Besides, some aspiration dialects show [s] instead of [h] for /s/ in resyllabification contexts, and this bleeding order between resyllabification and aspiration is interpreted to show that this /s/ becomes syllable onset after resyllabification (Kaisse, 1999).

An additional argument finds support in the production of Spanish tap /ɾ/, which is made up of two tongue gestures, one with the body of the tongue and one with the tip, as shown by Proctor (2009) for a wide variety of Spanish dialects. Moreover, in Castilian Spanish, when this phoneme appears in a resyllabification context, it never surfaces as [r], [ɹ], or [r̝]-which are some of the possible coda allophones-but only as the typical onset allophone [ɾ] (Blecua, 2001), which can be interpreted as the tongue-tip gesture being coordinated in phase with the following vowel.

Given these facts attested for [s] and [ɾ], we may invoke a parsimony principle and suppose that other Castilian Spanish consonants also undergo resyllabification when followed by vowel-initial words, although this would be no more than an extrapolation based on theoretical grounds. Based on our own data, in section 4.2 we suggest that the duration of flanking vowels may be interpreted as additional evidence of combined coda and onset properties also in the case of resyllabified /l/.

Besides, there are data of other languages that suggest that this could be a common phenomenon. For example, in an EMMA study for Korean, the labial gesture of derived onset /m/ was in-phase coordinated with the following vowel, i.e. it presented the same pattern of articulation as canonical onset /m/ (Cho et al., 2014).

This converging evidence of the maintenance of coda properties and the manifestation of onset properties points to an ambisyllabic nature of derived onsets, which results from their specific configuration of articulatory gestures. In the lexical form, the gestures of the consonant are coordinated to the previous vowel by virtue of an antiphase relation, as is typical of codas (Browman and Goldstein, 1992). Besides, when the consonant occurs before a vowel-initial word, there is a post-lexical restructuring that establishes an additional link with the following vowel. As a matter of fact, derived onsets can be interpreted as constituting an intermediate category between canonical codas and canonical onsets. This is suggested by the intermediate temporal duration observed for derived onsets (Strycharczuk and Kohlberger, 2016) and by their articulatory gestures, which are simultaneously phased to its two flanking vowels (Gick, 2003) ¹ .

An anonymous reviewer has suggested that the durational differences between derived and canonical onsets might be due to other reasons. For example, canonical onsets might be longer by virtue of the initial strengthening affecting consonants that are word-initial at the lexical level. While it is true that this effect on consonantal duration could be explained invoking other reasons, we believe that the ambisyllabic account has the advantage of also explaining the changes in duration of the flanking vowels, as will be discussed in the next section.

The duration of the surrounding vowels was also affected by the process of resyllabification in our results, but in a different manner depending on the identity of the consonant. On the one hand, a shorter duration of V1 and a longer duration of V2 were observed when in the presence of a derived onset /s/. On the other hand, V1 duration was not significantly affected by condition in the case of /n/ and /l/, and V2 was longer in the case of derived onset /n/ but not significantly different in the case of derived onset /l/.

These results are in contrast with the absence of significantly different durations of contextual vowels in the case of /s/ in the study by Strycharczuk and Kohlberger. In this sense, we wonder whether methodological differences in the production of the speech material might be responsible for such a contradiction. They included a single carrier sentence that was systematically repeated by participants, who might have emphasized or somehow highlighted the target sequence. Furthermore, it was probably easier for participants to realize the purpose of the researchers’ study, whereas the difference in our results might have arisen due to the naïve condition of our subjects.

Our interpretation of the durational differences for the contextual vowels found in this study is based on the different articulatory gestures involved in the production of each target consonant. More precisely, the fricative /s/ is achieved by a critical narrowing made with the tongue tip at the region of the alveolar ridge. The nasal /n/ also consists of a tongue-tip gesture, although in this case it requires a full contact with the alveolar ridge; and additionally, this is accompanied by a velum lowering gesture ² . In its turn, Spanish /l/ has been described as being made up of a tongue-tip gesture (with an alveolar contact and a lateral approximation) plus a tongue-body gesture realized as a pharyngeal approximation (Proctor, 2009) ³ .

If we suppose that the tongue-tip gesture of derived onset /s/ is simultaneously linked in antiphase to V1 and in phase to V2, a compromise must be met between these two opposing demands. In other words, the beginning of derived onsets cannot be fully aligned with the beginning of V2, as would be expected of a canonical onset by virtue of their in-phase coupling, but neither does it coincide with the halfpoint of V1 gestures, as would correspond to a coda: it rather lies in the middle of those points (Figure 3). This could explain an earlier timing of the consonant gesture in comparison with that of canonical onsets, thus resulting in a larger overlap with V1-which is therefore shortened-and a lesser overlap with V2-which is therefore lengthened.

Similarly for /n/, the ambisyllabic character of the lingual gesture might explain the lengthening observed for V2. However, a similar shortening of V1 as the one for /s/ was not found for /n/ (nor for /l/). A possible reason for this could be related to the specific constriction degree involved in the realization of these consonants. For example, while /s/ and /n/ are clearly a continuant and an interrupted consonant, respectively, /l/ has been described to behave differently according to the continuant/interrupted value depending on the language (Mielke, 2005). In Spanish, the evidence suggests that /l/ behaves as an interrupted consonant at least from the phonological point of view, as can be concluded from the fact that it patterns together with other interrupted consonants in the allophonic rule that predicts a voiced stop (instead of a spirant) after a homorganic interrupted (Lahoz-Bengoechea, 2015, § 4.3.1; Lozano, 2000 [1978], p. 325).

Thus, nasals and laterals require a full articulatory contact, similarly to stops and differently from fricatives. It is known that, as compared to fricatives, stops show a higher velocity in the articulatory gesture (Löfqvist and Gracco, 1997), and perhaps it could be interpreted that the same applies to nasals and laterals.

The faster achievement of the target and the subsequent shorter duration of the gesture might possibly explain why, even if the beginning of that gesture is slightly anticipated with respect to the canonical condition, the consonant barely overlaps with the preceding vowel, and consequently the duration of V1 is not significantly affected by the derived condition. In any case, such a possible difference in the articulatory velocity needs to be further studied using ultrasound or articulographic measures. Nonetheless, this behavior might be based on another underlying property affecting both consonants.

Finally, in the case of coda /l/ the consonant is typically made up of a tongue-body gesture (antiphased to V1) and a tongue-tip gesture (antiphased to the tongue-body gesture), which has been also confirmed for Spanish (Proctor, 2009). By virtue of resyllabification and based on an ambisyllabic account, the tongue-tip gesture would acquire an additional link to V2 (in phase). This predicts a simultaneous beginning of both the tongue-tip and the V2 gestures and a V2 similar in duration to the one following a canonical onset.

In summary, our results suggest, firstly, that consonants in the derived onset condition retain coda properties as shown by their shorter duration with respect to canonical onsets. Secondly, the onset status of resyllabified consonants can be attested by comparing the different durational behavior of the contextual vowels depending on the consonant identity. More precisely, both an ambisyllabic and a canonical coda interpretation would predict a shorter duration of V1 and a longer duration of V2 with respect to canonical onsets, and this is exactly what was observed for /s/. In both accounts, the antiphase coordination with V1 results in a gestural overlap with this preceding segment, which therefore undergoes a temporal reduction. In the case of V2, the greatest overlap would occur with a canonical onset, while an ambisyllabic consonant would show an intermediate overlap due to the opposing demands of the coordination with both flanking vowels, and a canonical coda would not show any overlap at all. Consequently, both the ambisyllabic and the canonical coda accounts would predict a longer duration of V2 than that of canonical onsets. By the same token, the observed increase in duration of V2 for resyllabified /n/ could be equally explained by both accounts just mentioned (Figure 4).

Our results for /l/, however, allow to disambiguate between those two possible interpretations, inasmuch as a canonical coda-as just described-would predict a longer V2 than in the case of canonical onsets, whereas the in-phase link invoked by the ambisyllabic analysis does predict a V2 similar in duration to that of canonical onsets, precisely as can be seen in our results. (Figure 4).

The possibility of an ambisyllabic representation for derived onsets was considered by Strycharczuk and Kohlberger (2016) in the discussion of their results, although they objected that two different phonological objects such as derived onsets and geminate consonants cannot have the same representation, and they reserved ambisyllabicity for the latter, especially since geminates have traditionally been analyzed in autosegmental terms as occupying two prosodic slots, i.e. coda and onset. On the contrary, we would like to suggest that geminates can be given a different phonological representation, while derived onsets may actually be analyzed as ambisyllabic. Therefore, in the following we offer evidence against an ambisyllabic representation of geminates, which in our view cannot constitute one single segment simultaneously associated to both the coda and the onset positions.

The first piece of evidence pointing to a possible reinterpretation of geminates can be found in Cypriot Greek, where it has been observed that f₀ minima are approximately aligned with an intermediate point of the geminate consonant, so that the observed pitch rise takes place in its second half (Tserdanelis and Arvaniti, 2001). A similar tonal alignment has also been reported for concatenated geminates in English (Gao and Xu, 2010). As a result, it is possible to infer that geminates are not one single but longer consonant, since otherwise tonal events would be expected to align rather with its beginning, as they do in the case of singletons.

In support of this view, we draw on the framework of articulatory phonology, which initially posited the notion of the articulatory gesture as the basic unit of phonological events (Browman and Goldstein, 1992), as previously mentioned. The model of articulatory phonology was later enriched with two more types of gestures: on the one hand, gestures aligned with the boundaries of prosodic constituents affecting the dynamics of articulation-i.e. π-gestures, (Byrd and Saltzman, 2003)-and on the other hand, gestures targeting a H(igh) tone and L(ow) tone-i.e. tonal gestures (Gao, 2008). The dynamics and overlap of these three different types of gestures are determined by their coordination-whether in phase or antiphase-in the gestural score of articulation, where the beginning of each gesture serves as the anchor point in the control of coordination (see Hall, 2017 for a review).

Specifically, the beginning of tonal gestures is coordinated in phase with the beginning of the gestures belonging to the nuclear vowel. Onset consonants are also coupled in phase with the nuclear vowel, so the onset is estimated to begin roughly in alignment with the tonal gesture. Since Cypriot Greek shows a pitch turning point at about the middle of the geminate consonant, this point should be interpreted as the articulatory beginning of the vowel and also as the beginning of the onset. Crucially, the geminate cannot be considered one single but longer articulatory gesture-if that were the case, the vowel would start halfway through the consonant, in what could be seen as a sort of antiphase relation. As just said, anchor points for coordination are necessarily located at the beginning of gestures. Since this is the case, and the onset is required to be in phase with the vowel, the f₀ turning point reveals the start of a new gesture, leading to the conclusion that geminates must be represented as a sequence of two consonants.

Secondly, our interpretation of geminates as sequences of two identical segments is supported by Tilsen’s description of the acquisition of phonological patterns in speech development (Tilsen, 2016). In his account, Tilsen points out that geminate consonants arise by reselecting the same motor plan as that of the previously realized segment. In the first stage of development, children have not learned yet how to reselect action units and therefore produce geminates as short as singletons. On a subsequent stage, they can reselect the motor plan, but rely solely on external feedback to know that the first instance of that plan is completed prior to reselecting it. In other words, not until they hear the acoustic consequences of the first unit or feel the appropriate degree of muscular tension do they proceed to reselect the motor plan. This gives rise to an abnormally long geminate production. Finally at a later stage of the development, children learn to anticipate the consequences of their articulation, and this anticipation-called internalized feedback-, allows them to reselect units in advance, so that the greater overlap between consecutive actions yields geminates as long as expected.

Consequently, this view allows to posit a similar account for lexical geminates and for geminates concatenated across word boundaries, both consisting of a sequence of two consonants. The difference is that lexical geminates are produced by reselecting one unit of motor planning, whereas concatenated geminates correspond to two such units, albeit identical to each other. Such articulatory difference may account for other well-known phonetic and phonological differences between lexical and concatenated geminates.

A third piece of evidence in favor of the rearticulation of geminates is provided by Miller (1987) for Levantine Arabic, where spikes of energy corresponding to the subtle release bursts have been observed in the middle of geminate consonants on certain occasions. In line with Tilsen’s view (2016), this is suggestive of the reselection of a previously executed motor plan, which results in a new constriction gesture to prolong the consonant. Following Tilsen’s account, it is possible to interpret the appearance of such subtle bursts as arising from an inefficient internal feedback. Thus, the speaker must rely rather on the external feedback-i.e. the acoustic effect provided by the release-to detect when the gestural target has been reached prior to reselecting the motor commands that result in a geminate consonant.

Strycharczuk and Kohlberger (2016) discarded an ambisyllabic representation for derived onsets because they argued that such representation must be given to geminates. Our point is not to extend the data from the mentioned languages--Cypriot Greek or Levantine Arabic--to Spanish, but rather to show that ambisyllabicity is not a good analysis for geminates in a wide array of languages. Therefore, we argue that an ambisyllabic analysis remains available as a possible representation for derived onsets.

For all these reasons and as previously mentioned, there is sufficient evidence to interpret resyllabified consonants as onsets after the post-lexical reorganization of syllable structure occurring in resyllabification (see section 4.2). In addition to this, it is possible to consider that derived onsets still show coda properties (see sections 4.1 and 4.2) and have an intermediate status that points to an ambisyllabic representation (Figure 5).