Examining Oral (Dis)Fluency in—uh– —Spanish as a Heritage Language

Abstract

Silence, self-interruptions, or hesitations in spontaneous speech are often interpreted as markers of oral disfluency as they lead to difficulties in delivering a message and in processing it. The main purpose of this study is to examine how such markers of discourse structure factor into the overall oral fluency of 58 US Spanish heritage language learners enrolled in Spanish classes at the college level. Participants were grouped according to age of onset of bilingualism (i.e., sequential or simultaneous) and the order in which they acquired each language (i.e., English first or Spanish first). After completing a semi-controlled oral production task, in both Spanish and English, breakdown pauses and repair pauses were extracted and then analyzed in terms of quantity, quality, and mean duration. Our findings revealed (i) that all groups produced shorter pauses in English, their dominant language; and (ii) that all experimental groups behaved very similarly in Spanish despite having had different experiences with bilingualism growing up. Albeit tentatively, given the sample size and the nature of the present study, we take these findings to suggest that type of heritage bilingualism and the order in which each language was acquired does not seem to play a significant role in terms of production of breakdown and repair pauses.

1. Introduction

Pauses, hesitations, self-interruptions, and reformulations are all part of everyday communication and play a very important role in the process of meaning–making in speech acts (e.g., Grosjean 1980). From the speaker’s point of view, a pause can mean that they may be planning what to say next, may be having difficulties finding the right words, or may be bringing attention to their message (e.g., Cenoz 2000). From the listeners’ perspective, pauses are necessary for information processing purposes. However, an increase in pausing could cause listeners’ comprehension to decline as intelligibility lessens (e.g., O’Connell and Kowal 2005; Krause and Braida 2002). Although such markers of speech disfluency have been widely investigated in the fields of first (L1)1 and second language (L2) acquisition (e.g., de Jong 2016; Cenoz 2000; García-Amaya 2009), to our knowledge, this phenomenon has not received the same attention among Spanish heritage language learners (HLLs) in the context of the United States (see Erker and Bruso 2017). Additionally, given that the speech of heritage speakers is often described as having what some refer to as a slight “heritage accent” (e.g., Polinsky and Kagan 2007; Polinsky and Scontras 2019; Rao 2014; Shin 2005), and considering the significant role that pauses and repairs play in oral communication, with this study, we seek to determine whether these markers of oral disfluency contribute to the perception of said accent in Spanish as a heritage language (HL). This, we believe, is an important and necessary first step towards developing an understanding of what constitutes speech disfluencies in asymmetrical bilingual environments. It is well known by now that HLLs make up a very complex and heterogeneous group, as they experience different trajectories in terms of age of onset of bilingualism, exposure, and opportunities to use the HL, access to education in the HL, as well as other contextual factors such as social stigma and prejudice (e.g., Montrul and Polinsky 2021; Montrul 2023; López et al. 2023; López 2020; Pascual y Cabo and Montrul 2021). Because of these unique trajectories, the nature of their bilingual reality should not be understood as a binary, but as a continuum rich in possibilities and outcomes.)

Notably, previous work in this area has focused almost exclusively on examining the linguistic systems of heritage speakers who acquired both languages either simultaneously or sequentially, starting always with the home/heritage language. In fact, dare we say, this is the stereotypical linguistic profile that usually comes to mind when we think of heritage speakers. And, while in this study, we present data from participants that can be categorized in such a way, we also document the linguistic practices of an understudied subset of heritage speakers that do not follow the traditional “home/heritage language first” condition that appears in most definitions (e.g., Valdés 2001; Benmamoun et al. 2010). That is, part of the data to be analyzed come from participants who were exposed to the HL sequentially after having first acquired the dominant language in society.

Although these bilinguals have often been excluded from the traditional conceptualization of heritage speakers, this—we believe—will prove to be an interesting contribution, as more information on the differences and similarities between the HL groups can provide further insight in the research concerning age of onset of bilingualism in bilingual production/fluency.

Our interest in examining HL oral (dis)fluency stems from the general assumptions made about HLLs’ oral skills: early exposure to naturalistic input in the HL affords HLLs with advantages in production and perception compared to traditional L2 learners (e.g., Polinsky and Kagan 2007; Au et al. 2008). We focus on the quantity and duration of gaps between words rather than the sounds themselves as a way to analyze and better understand oral (dis)fluency markers in the speech of HLLs. Previous research with other bilinguals (i.e., traditional L2 learners) has revealed significant instabilities regarding breakdowns and repairs compared to monolingual speakers of the same language (e.g., Wiese 1984; de Jong 2016). In fact, slow speech production rates and increased number of pauses as well as repairs such as repetitions, self-corrections, and reformulations are typically understood as L2-specific phenomena, likely due to a combination of conceptualization and formulation demands (Duran-Karaoz and Tavakoli 2020). As for HLLs, it is largely unknown whether the processes are automatized and if similar fluency vulnerabilities are encountered (Tavakoli 2020). Considering this, the question of whether they conform closer to L1 or L2 fluency patterns is, at the very least, an intriguing one. To find answers to this question, in this study, we seek to examine and describe how pauses and repairs factor into the overall oral fluency of Spanish HLLs.

2. Understanding Pauses and Other Markers of Oral (Dis)Fluency

As mentioned earlier, pauses, reformulations, hesitations, or even silence are commonplace in everyday speech. Though apparently basic and meaningless in their ubiquitousness, these subtleties constitute complex communicative phenomena that play an important role in what we are broadly referring to as oral (dis)fluency: the oral outputs which make oral productions disfluent (e.g., Gao and Du 2013). Given that pauses and hesitations are a critical part of speech community and conversational norms, and because native speakers are generally considered to make the best use of time constraints when producing an uninterrupted stream of smooth and hesitation-free speech (e.g., Moreno Fernández 2002), achieving “native-like” or “monolingual-like” oral fluency often becomes one of the primary goals for most L2 learners. Although HLLs can be considered “native” speakers of the HL (e.g., Kupisch and Rothman 2016), because they are often judged on their pronunciation or on how fluent their speech production sounds in their HL (e.g., “pocho Spanish”2, Sánchez-Muñoz 2016; Tseng 2021), many of them seek to sound like “monolingual” speakers too. As a result of pressures to meet the HL community linguistic expectations, HLLs may not only be negatively affected in their linguistic identity and their self-efficacy beliefs, but may also develop HL anxiety and HL shyness (e.g., Prada et al. 2020; Ortega 2021; Tseng 2021). This may make them feel insecure about their speaking skills, and in turn, speaking less may affect fluency and pauses. Early explanations of the hesitation differences (pausing behavior) in bilinguals were based on the degree of automatization across languages (Wiese 1984). According to Wiese, when bilingual speakers use their non-dominant language, their speech production is less automatic, which often results in increased planning time and self-corrections (Wiese 1984).

More recently, Segalowitz (2010) identified vulnerability points in which bilingual individuals may encounter speech formulation and production difficulties: when drawing on the language-specific information (e.g., syntactic organization, morphological encoding), particular challenges arise. These challenges are reflected temporally in fluency with the relationship between speed and pause. Although many studies have utilized speech rate (in syllables per time unit), speed of delivery is not the only unit of analysis for fluency. Disfluency markers interrupt the stream of words without contributing propositional content to the utterance (Tree 1995); hence, pause time is considered an adequate measure for fluency measure as well. Pause classifications have typically been categorized with disfluency markers as follows:

• Pauses; silent, empty, or unfilled pauses: empty interruptions in speech. The threshold of duration ranges between 100 ms (Riazantseva 2001), 250 ms (Bosker et al. 2013; Leonard and Shea 2017), and 400 ms (Derwing et al. 2004) (e.g., “There was a bunny and a (-) big dog”).
• Filled pauses or fillers: interruptions filled with sounds like uh or er, a nasal consonant alone (e.g., mm), or a vowel followed by a nasal consonant (e.g., em, um) (e.g., “The (hum) wolf was ugly”).
• Repetitions: repetitions of syllables or words, unless duplications are semantically motivated, made to provide clarification or specification (e.g., “The thing is that (-) the thing is that the big dog never learns”.
• Self-corrections: modification of the original speech before interruption because the sentence material (generally grammar), in eyes (or voice) of the speaker, needs rectification (e.g., and then the wolf catched (-) caught the rabbit”).

Pauses one and two are considered “covert” repairs or breakdown fluency, because the speaker has little information to assess what may have gone wrong during speech production to explicitly fix it. Conversely, pauses three and four are “overt” (Levelt 1983) or simply “repair” fluency, because the speaker rejects their speech as they were able to infer what went wrong.

Oral Fluency in Bilingual Literature

Because oral fluency is an essential characteristic for successful communication, this is an area that has been widely researched in the field of L2 acquisition (e.g., Cenoz 2000; Williams and Korko 2019; de Jong 2016). Space limitations will allow reference to this, but a few relevant studies in Spanish for L2 Spanish and HLL populations to show how experience in language learning affects fluency development. For example, García-Amaya (2009, 2015) investigated the dimensions of fluency in different learning environments (at home vs. study abroad) and analyzed both speech rate measures (e.g., total spoken words, total syllables) and breakdown and repair measures (total filled pauses, total repetitions). When looking at speech rate measures, the study abroad group produced the highest range of spoken words and more syllables, and spoke at a faster rate than the other learner groups. There was an overlap among the learner groups in the repetition and repair analysis which did not provide a reliable measure to show fluency gains. However, the fastest rate of speech with an increase in short pause use may serve as processing mechanisms to mitigate the greater demands of lexical and syntactic planning of communicating in the study abroad context.

More recently, García-Amaya (2020) researched a group of Afrikaans migrant speakers in Patagonia, Argentina. His interest was to look at fluency measures (syllable length and sentence length), breakdown measures (silent and filled pauses per minute), and repair fluency (number of reformulations and self-corrections) in the dominant language (Spanish) of a group whose L1 (Afrikaans) was under attrition. He collected data from a total of 24 participants (both Afrikaans bilinguals and Spanish monolinguals) via personal interviews and a background questionnaire and extracted information to describe participants’ fluency in the dominant language. When comparing Spanish monolinguals and Afrikaans–Spanish bilinguals, both obtained similar results for production speed. However, the bilingual participants produced more silent pauses, more reformulations, and fewer fillers. In general terms, these findings were interpreted as markers of disfluent speech because the speakers would have used these elements to process and plan their discourse. A further analysis of Afrikaans–Spanish filled pauses (García-Amaya and Lang 2020) revealed that, compared to monolinguals, bilinguals display separate F1 and F2 vowel values in both L1 and L2 but no consonant differences, which may indicate that cross-language phonetic inference is affected by the same patterns as those in lexical items. As a whole, García-Amaya’s research in Patagonia indicates that fluency analysis is not only an interest for foreign language learning, but for other bilingual communities as well. Additionally, this study reminds researchers to investigate both the dominant and non-dominant languages in order to account for long-term effects of bilingualism on fluency.

Compared to the information that deals with L2 pause production, less is known about HLLs. The lack of work in this area may be largely due to the anecdotal nature of the evidence reported early on, indicating that HLLs’ pronunciation was practically monolingual-like3. Compared to other linguistic domains, HLLs may have more advantages in oral competence (e.g., Benmamoun et al. 2010). However, this does not mean that HLLs’ sound system is not vulnerable or impervious to crosslinguistic influence. Acknowledging that oral disfluencies may be a result of crosslinguistic influence, the last 10 years have witnessed a surge in research on HL speech production (e.g., Rao and Ronquest 2015; Rao 2016; Colantoni et al. 2016). This research has shown that crosslinguistic influence may in fact be responsible for many HL sound patterns. For instance, vowels show a reduction when placed in an unstressed syllable, and this effect is intensified in spontaneous speech (e.g., Ronquest 2013, 2016). The study of consonants has shown that, in general, HLLs resemble speech closer to native monolinguals than L2 learners (Knightly et al. 2003; Au et al. 2002); especially, earlier exposure implied a more similar pronunciation to L1 production (Amengual 2019), but certain phonemes are more vulnerable to English influence, for example, [β], which is an allophone of /b/ (Rao 2014, 2015).

Although the study of oral disfluencies in HL discourse production is practically non-existent, there are, however, a few important exceptions. For instance, Erker and Bruso (2017) examined the possible effects of language contact between Spanish and English on filler4 behavior (e.g., eh, hum, and er). Their main goal was to analyze acoustically the number of English and Spanish fillers, which are language-specific and become a locus of cross-linguistic contrast. Their secondary goal was to investigate convergence as a result of language contact. That is, if two languages become more structurally similar over time through the adoption of features from one language (usually the dominant, in this case, English) into the other (usually subordinate, here, Spanish). Informants completed a self-reported proficiency and language use survey and participated in an interview. The results revealed a modest overall preference for lexical fillers in the data; “y” (and) was the most preferred among speakers. There was a clear relationship between a higher degree of contact and preference for English fillers (hum, ah(m)), and those with a lower degree of contact preferred eh (m). As the contact intensity increased, speakers shifted towards an increase in more centralized vowels, but not quite like schwa. Erker and Bruso (2017) found a clear relationship between contact intensity and variation in phonological fillers, which was supported by the correlation of pause behavior with the intensity of the contact. In the vowels, a greater degree of contact corresponded to increasing centralized vowels in fillers. This was interpreted as evidence of contact-induced change.

In a related study, Enríquez (2020) examined pauses in dialogues between HLLs, L2 Spanish learners, and speakers in a Spanish-dominant country. Enríquez was interested in pauses as a politeness resource during oral interaction, since pauses may not only signal processing proficiency or cognitive load, but also pragmatic demand. Her participants performed three roleplays in English and in Spanish; from these, filled and silent pauses were quantified. After codifying the recordings, it was found that more fillers than silent pauses were used by HLLs, and vocalic elongations depended on proficiency. The interpretation was that silent pauses may generate anxiety, as speakers in a conversation tend to avoid silence and produce tense situations (Havekate 1987; Ragsdale 1976; Kasl and Mahl 1965).

3. The Present Study: Toward a Better Understanding of Pauses in Spanish Heritage Speech

With this study, we seek to describe and achieve a better understanding of HLLs’ (dis)fluency breakdowns in Spanish, their weaker language, by examining the effect of conceptual planning and the effect of language dominance on breakdown and repair fluency. Considering this, the research question that guided this study is the following:

What are the main characteristics of Spanish heritage speakers’ oral narratives in terms of metrics of breakdown fluency (filled and silent pauses) and repair fluency (reformulations and repetitions, self-corrections)?

Based on previous related work (e.g., García-Amaya 2009, 2015, 2020; Enríquez 2020), we can hypothesize that our participants will make more and longer silent and filled pauses in Spanish, due to the taxing nature of conceptual planning. Additionally, we anticipate that language dominance as a result of age of onset of bilingualism will also have an effect in the number of self-corrections, reformulations, and repetitions they make.

4. Methodology

4.1. Participants

A total of 58 heritage speakers (20 male, 38 female), all students of the Spanish Heritage Language (SHL) program at a large public research university, participated in this study. Their responses to an adapted version5 of the Bilingual Language Profile (BLP; Birdsong et al. 2012) provided us with biographical data (age range: 18–23, mean 19.8, SD = 2.653), language history, and current language use. All participants were university-age students and, at the time of data collection, they all resided in [STATE]. Students’ backgrounds varied, but given the general demographics of the university, they were mainly Cubans, Puerto Ricans, and Colombians of various sociolinguistic generations. They were all raised bilingually in the United States from early on, some with exposure to both Spanish and English from birth (n = 24), while others experienced the exposure sequentially (n = 34). Although they all received formal education in English growing up, at the time of this study, they were enrolled in Spanish HL classes in three different levels (intermediate, intermediate–advanced, advanced).

As we were looking into the biographical data of our participants, it became evident that within each class level, there were very different linguistic profiles. For one, the age of onset of bilingualism and the order in which they acquired each language varied considerably. Since timing of first exposure to English had been found to be a contributing factor to a number of linguistic outcomes (e.g., Montrul 2014; Pascual y Cabo and Gómez Soler 2015), we decided to group them by age of onset of bilingualism. Thus, participants were split into three groups: Simultaneous bilinguals (n = 24), English-first sequential bilinguals (n = 14), and Spanish-first sequential bilinguals6 (n = 20). As the group name indicates, simultaneous bilinguals are those who grew up with exposure to both languages from birth. Sequential bilinguals, on the other hand, are those who acquired the second language after having had sufficient exposure to master the first. For the English-first sequential participants, age of acquisition of Spanish was 6.5 (SD = 3.34), and for the Spanish-first group, age of acquisition of English was 4.5 (SD = 0.89). Normally, when we use the label sequential, we are referring to those HLLs who acquire the HL first; in our case, Spanish. Interestingly, however, in the present article, we examine data from a largely understudied group of HLLs: sequential speakers of Spanish whose first language was English. Looking at their linguistic background forms, those English-first sequential informants reported growing up with very occasional exposure to Spanish during the first years of their lives. Later on, however, they gained access to the HL through language programs in schools, which allowed them to develop productive language skills. In other words, had it not been for social-level influences that promoted HL development (Lynch 2000), they would have remained receptive bilinguals. With this study, we seek to contribute to examine this sample of bilinguals.

Table 1. Self-ratings of language proficiency in English and Spanish.

	Speak		Understand		Read		Write
	Eng	Spa	Eng	Spa	Eng	Spa	Eng	Spa
Seq_Eng first	6.0 (0)	3.64 (1.08)	5.93 (0.27)	4.5 (0.94)	5.9 (0.3)	4.07 (1.2)	6 (0)	3.57 (1.4)
Seq_Spa first	5.8 (0.52)	4.35 (0.9)	5.9 (0.44)	5.5 (0.61)	5.9 (0.31)	4.9 (0.79)	5.7 (0.73)	4.2 (1.06)
Simultaneous	5.8 (0.55)	3.7 (1.05)	6 (0)	5 (0.91)	5.95 (0.22)	4.58 (0.69)	5.79 (0.55)	3.7 (0.86)

1 = very low; 6 = very high. Eng = English, Spa = Spanish. Participant grouping: Seq_Spa first: sequential speakers; Spanish first acquired, Seq_Eng first: sequential speakers; English first acquired, Simultaneous: English and Spanish acquired at the same time.

below includes information about the participants’ perceived proficiency in reading, writing, speaking, and understanding in each language. As can be seen, our students see themselves as more fluent in English across the board. Importantly, all HLLs regardless of age of onset of bilingualism (i.e., all three groups), rated their skills in Spanish very similarly. Not surprisingly, the group that rated their Spanish proficiency the lowest was the sequential English-first group.

4.2. Instrument: Oral Production Task

Participants completed a semi-controlled production task in the university language computer studio. The task, based on the popular Russian cartoon Nu, pogodi! (Just wait!), consisted of (i) watching a 1 min videoclip in which a wolf was chasing a rabbit and trying to eat it, and (ii) describing the content of the video with a time limit of 3 min and without external help, dictionary access, or taking notes. They completed this task two times; first in Spanish and then in English (a different videoclip, similar storyline). Analyzing their recordings in both languages allowed us to have an intraspeaker comparative baseline. The use of elicited video narratives offers several advantages (e.g., Kopotev et al. 2020; Echevarría and Pascual y Cabo 2023). For example, not only has this particular task been used with success in previous research studies (e.g., Parra et al. 2018), but presenting the plot of the cartoon as a cue ensures a more objective comparison across subjects. Importantly, since the plot of the cartoon was fairly simple, even those students with lower proficiency skills could complete the task without many difficulties.

The narratives produced by the students were digitally recorded, transcribed, and independently coded for analysis by two researchers. Each recording was uploaded to Audacity as an MPEG Layer-3 audio file and pauses were initially identified aurally (breakdown or repair was heard) and subsequently confirmed visually (seeing a gap at the wave script). Then, the start and end of each pause was checked with the transcript and recorded in a separate Excel file. Following previous research, we decided to only extract silent and filled pauses that were greater than 250 ms and in a mid-clause location (Bosker et al. 2013; Leonard and Shea 2017), as pauses shorter than 250 ms are not regarded as hesitation phenomena (Riggenbach 1991). As a reference of fluency, Spanish monolingual speakers produce pause durations from 391 to 482 ms and an average of 8 per minute (Schwab 2015). However, no monolingual control group was assessed because the goal of the study was not to compare HLLs with monolingual measures but themselves. As a reference, we are using participants’ recordings in English (in results section as “English reference”, in plots, “English_ref”).

For the purpose of this study, we followed the typical categories for disfluency markers, which we described earlier in Section 2 (Götz 2013; Kormos 2006; Maclay and Osgood 1959; Levelt 1983). First, measures of breakdown fluency were divided into silent pauses (stretches of silence within the stream of speech) and fillers (lexical or non-lexical voiced utterances, such as I, erm, like, and actually, and prosodic markers such as laughter and sighs that interrupt the stream of speech). Second, measures of repair fluency were categorized as reformulations (abandonment of an utterance followed by its immediate revision to improve coherence, including false starts and repetitions) and self-corrections (attempted replacement of perceived non-standard output with a form that a fluent speaker would recognize as standard). It may not be resolved into a grammatical production, but it is noticed by the speaker as an “error”. The example below in (1) illustrates how we codified an excerpt of 17.42 s from a simultaneous bilingual:

(1)

El lobo [eh-0.709 (filled pause)] es un poquito gordito y el conejo muy [0.523 (silent pause)] po-pequeño y hermoso y el lobo quiero [0.430 (correction] quiere [0.261 (correction)] quería [0.444 (reformulation)] el lobo quería [0.846 (silent pause)] almorzar el conejo.

(The wolf is a bit chubby and the rabbit very small and handsome and the wolf want wants wanted the wolf wanted to eat the rabbit).

There were two silent pauses, one filled pause, two corrections, and one reformulation. Only breakdowns and repairs over 250 ms were counted; hence, the correction “po-pequeño”, smaller than that length, was not included. Participant recordings lasted from 1 to 3 min; hence, pause duration and quantity was averaged to a minute.

5. Results

Upon the examination of participants’ transcripts, we extracted the average pause production per minute of speech and average pause type duration per minute for each group. Subsequent data were submitted for statistical analysis on RStudio (RStudio Team 2020) using base R and effsize packages (Torchiano 2020). Given that traditional statistical tests and p values have been heavily criticized as not being the best proxy for gauging practical significance in the field of applied linguistics (e.g., Plonsky and Oswald 2014), we decided to conduct effect size analyses, which seem to better characterize the meaningfulness of group differences (e.g., Cohen 1992; Plonsky and Oswald 2014). And so, following general guidelines, an effect size of 0.20 or lower is considered a small effect size, 0.50 is a moderate effect size, and anything 0.80 or higher is considered a large effect size (Cohen 1992; but see Plonsky and Oswald 2014). For ease of exposition,

Table 2. Effect size calculations for the four pause types (average duration and length) across all groups.

	Simultaneous English First	Simultaneous Spanish First	English First Spanish First	Simultaneous English Reference	English First English Reference	Spanish First English Reference
Pauses
count	0.15	0.62	0.62	0.56	0.57	0.13
duration	0.24	0.04	0.23	1.49	1.57	1.04
Filled pauses
count	0.28	0.32	0.11	0.19	0.07	0.15
duration	0.35	0.69	0.53	0.82	0.74	0.06
Reformulations
count	0.76	0.65	1.4	0.29	0.95	0.21
duration	0.23	0.08	0.11	1.01	1.53	0.82
Corrections
count	0.41	0.69	0.41	1.16	1.27	1.02
duration	0.22	0.45	0.8	1.02	1.47	1.28

Effect size calculations for the four pause types (average duration and length) across all groups. Large effect sizes are in bold.

below displays the effect sizes for each analysis.

Figure 1 shows the mean number of silent pauses per minute and the mean duration of participants’ silent pauses. As can be seen, the group that made, on average, more pauses was English-first (mean = 8.9, SD = 5.59), closely followed by Simultaneous (mean = 8.8, SD = 3.81) and Spanish-first (mean = 6.5, SD = 4.07). In terms of pause duration, once again, English-first produced longer pauses (mean = 1.25, SD = 0.37), followed by Simultaneous (mean = 1.12, SD = 0.32) and Spanish-first (mean = 1.08, SD = 0.48). The English reference (participants’ recordings in English) for quantity of silent pauses was the closest to the Spanish-first results (mean 6.96, SD = 3.16), and the shortest in duration (mean = 0.716, SD = 0.2). Cohen’s d calculations yielded medium effect sizes in the count of pauses of Simultaneous and Spanish-first (d = 0.62) for Spanish-first and English-first (d = 0.62). The smallest effect size comparing the number of pauses produced was found between Simultaneous and English-first (d = 0.15) and Spanish-first and English reference (d = 0.13). The largest effect sizes were in the duration of silent pauses, especially between the English reference and the other three groups (simultaneous-English reference: d = 1.49; Spanish-first English reference: d = 1.57; English-first- English reference: d = 1.04).

Figure 2 includes the number and duration of filled pauses by group. Contrary to what was found for silent pauses, filled pauses were actually quite similar across both languages and groups, the English reference (English_ref), and the recordings in Spanish (other three groups). Participants were producing two to three pauses per minute and durations ranged from 0.388 to 0.576 s. These plots not only show that participants produced less filled than silent pauses but also that there were more outliers. Cohen’s d calculations yielded small effect sizes across all groups (largest d was 0.32 reported between Spanish-first and English-first) in the quantity of fillers and it was filler duration between Simultaneous and English reference (d = 0.82) and English reference and English-first (d = 0.74) that presented the largest effect sizes in this category of pauses.

Figure 3 displays the number and duration of reformulations produced by group. As can be seen, participants produced fewer examples of repair fluency pauses, but the length of reformulations was similar to that of silent pauses. The group who made, on average, more reformulations was English-first (mean = 3.1, SD = 2.63), followed by Simultaneous (mean = 2.2, SD = 1.94). Spanish-first (numbers) was the closest to the reference of English (English_ref) (mean = 6.5, SD = 4.07). For duration, Simultaneous produced longer pauses (mean = 1.25, SD = 0.37), followed by English-first (mean = 1.12, SD = 0.32). Spanish-first (mean = 1.08, SD = 0.48) was below one second, but English reference was the fastest (numbers). There were more differences among groups in this area. Large effect sizes were obtained for number of reformulations between Simultaneous and Spanish-first (d = 0.76), Spanish-first and English-first (d = 1.4), and English reference and English-first; the other group combinations yielded medium effect sizes (d = 0.47). The group contrast for duration yielded large effect sizes in comparing the English reference with Spanish-first (d = 0.82), English-first (d = 1.53), and Simultaneous (d = 1.01).

Figure 4 shows the quantity and duration of corrections made by group. As can be seen, English-first once again produced more corrections compared to Spanish-first (and the same compared to Simultaneous). However, it should be noted that on average they only produce one correction every minute of speech. There were no examples of corrections made by participants within the English reference group (English_ref), because English was the dominant language. Effect sizes were only obtained when comparing the three groups. For duration, medium effect sizes were found between Spanish-first and English-first (d = 0.41), Simultaneous and Spanish-first (d = 0.69), and Simultaneous and English-first (d = 0.41). Lastly, the duration of such corrections also yielded the same pattern of results with medium-to-moderate size effects: Simultaneous and Spanish-first (d = 0.45), Spanish-first and English-first (d = 0.8), and Simultaneous and English-first (d = 0.22).

6. Discussion and Conclusions

The main goal of this study was to identify, document, and analyze pauses and repairs as markers of oral (dis)fluency in Spanish HL in the context of the United States. Given the role such (dis)fluency markers play in everyday communication, this constitutes a significant and necessary first step towards developing an understanding of speech disfluencies in asymmetrical bilingual environments. Specifically, the question that guided our study was: What are the main characteristics of Spanish heritage speakers’ oral narratives in terms of markers of breakdown fluency (i.e., filled and silent pauses) and repair fluency (i.e., reformulations and repetitions, self-corrections) in both English and Spanish?

Generally speaking, due to the taxing nature associated with conceptual planning (e.g., Felker et al. 2019), we predicted that our informants would not only produce more and longer pauses in their non-dominant language (Spanish) than in their dominant language (English), but that they would also produce more self-corrections and repetitions. Although we mostly found low-to-moderate effect sizes, the trends observed generally support these two premises given that the values recorded for both categories were consistently higher for Spanish than for English.

When examining the English measures across groups, we observed that their pauses were very short in comparison to those produced in Spanish. Furthermore, they produced very few reformulations and not one single correction. As a whole, none of this is comes as a surprise given that, as mentioned earlier, all informants reported English to be their dominant language.

With regard to the Spanish measures of breakdown fluency, within- and across-group comparisons revealed an interesting trend: despite finding some differences, our groups behaved very similarly, sometimes producing a range of overlap values. This is interesting given that each of these three groups were distinct in terms of their age of onset of bilingualism and the order in which they acquired each language. To be sure, participants in the Spanish-first group produced fewer and shorter pauses overall compared to simultaneous and English-first bilinguals, but Cohen’s d calculations yielded small and moderate effect sizes. We take this to indicate that, at least for the particular conditions and participants examined in this study, age of onset of bilingualism does not seem to be a factor. Although we do not have data to determine it with certainty, it may be the case that language competition is affecting HLLs’ speed and accuracy of lexical processing similarly. Future studies should assess this observation.

Furthermore, if we compare the average bilingual values obtained herein (avg. 850 ms for silent pauses) with those reported in previous studies for monolingual Spanish speakers (range 400 ms–700 ms, e.g., Blondet 2006; 541 ms, Cremades 2016; 521 ms, Llisterri et al. 2019) or for Spanish-dominant bilinguals (avg. 685 ms, e.g., García Amaya 2020), we see that Spanish-dominant speakers generally resolve pauses slightly faster than English-dominant ones tested in this study. Although not conclusively, given the methodological and sampling differences that exist among our study and the comparable cited studies (e.g., interviews vs. narration task), in light of the data available, it seems reasonable to suggest that our HLLs’ underlying performances might be related to language dominance. Future studies should explore these possibilities.

Second, looking separately at each pause type, we did observe that our informants produced more occurrences of breakdown fluency than of repair fluency. And while previous related research found that Spanish HLLs produced more filled pauses than silent pauses (Enríquez 2020), our results reveal the opposite trend. Enríquez (2020) attributed the larger number of filled pauses in HLLs’ dialogues to silent pauses being interpreted as a result of (and contributing to) stress in conversation. Given that our participants’ data were collected using a narration task and not in natural conversation, it is possible that filled pauses were not produced in the same ratio, essentially because the task was less stressful than a face-to-face situation. Furthermore, in a natural and spontaneous conversation, speakers depend on someone else’s interaction, whereas in a narration task, this other person is not present, which may reduce pauses. They did rely on silent pauses, which are known to convey cognitive load and are used as a strategy to retrieve lexical knowledge (e.g., Cenoz 2000).

When examining our participants’ metrics of repair fluency, it should be noted that, as a whole, they produced very few reformulations and self-corrections. This is as intriguing as it was unexpected, particularly because we found no differences within or across groups and between their production in English and Spanish. On the one hand, one could argue that their English production needed very few corrections and reformulations because they are all English-dominant. For their Spanish corrections, even though they did not always correct themselves, at times they made mismatches with very simple structures such as such number/gender agreement. Although we do not have conclusive data to answer this question, it is possible that the low effect sizes between groups were the result of a task effect. That is, in English, the task was too simple for them, so they made no errors that needed correction or reformulation (i.e., no measures in English recordings or one per minute). Although in Spanish they did make mistakes, the low-stress nature of the task did not trigger the participants to monitor their own linguistic output and, therefore, they ended up not correcting themselves.

Despite contributing to our understanding of Spanish HL speech production, this study is not without limitations. For one, the number of participants was rather low, so the results should be interpreted with caution. Additionally, as expected in a university setting, our participants represented different linguistic backgrounds with different degrees of proficiency. It would be interesting to see if results would differ if dialectal variation were controlled. For instance, in certain varieties (e.g., Mexico and Puerto Rico), fillers may not be isolated vowels, but lexical fillers elongations are preferred to the use of isolated fillers (e.g., Torres and Potowski 2008; Said-Mohand 2008). Furthermore, as pointed out, the nature of the production task employed here seems to have influenced our findings. A limitation was the fact that each recording was under three minutes, based on Parra et al.’s (2018) study design. Another shortcoming was the lack of counterbalanced order: participants provided their recordings in Spanish first, and then English. Future studies should attempt to obtain data not only from controlled settings, such as reporting narrations, but also interviews in different contexts (formal dialogues, colloquial conversation). By doing so, we will be able to see how disfluencies are affected by the choice of methods.