Highlights

• • Unimodal bilinguals process bilingual audio (speech)–visual (written) simultaneously

• • Integration of bilingual audio (speech)–visual (written) stimuli is automatic

• • Effects in processing are driven by modality and not by language dominance

2. Experiment 1: Italian (L1) stimuli with English (L2) subtitles and blurred mouth

In our first experiment, participants were Italian (L1)–English (L2) bilinguals and were presented with short video clips showing a person uttering a word. The spoken language was the L1 (Italian) and the written language was the L2 (English). The two unimodal conditions were as follows: i) Spoken L1 and ii) Written L2. The bimodal condition was Spoken L1 + Written L2.

In the video clips, the mouth of the speaker was blurred. We took this decision as we did not want any L1 linguistic interference in the muted stimuli showing the L2 written word, and it is well known that watching lip movements contributes to speech perception (e.g., Grant & Walden, Reference Grant and Walden1996; McGurk & MacDonald, Reference McGurk and MacDonald1976; Peelle & Sommers, Reference Peelle and Sommers2015).

Our research questions were as follows: Is bimodal bilingual simultaneous lexical access costly for unimodal bilinguals? If it is costly due to the difficulty in processing two languages at the same time, we would expect longer reaction times in the bimodal condition compared to the unimodal conditions. On the contrary, if double lexical access is an advantage, we should expect faster reaction times in the bimodal condition compared to both unimodal conditions. Moreover, what is the impact of bilingualism on the processing advantage of written versus auditory stimuli? Being the written language participants’ L2, it might be that participants rely more on their L1, thus lowering any advantage of the written modality over the spoken one. In this case, any bimodality effects could be overridden. This scenario is plausible, also considering that participants in the present study are unbalanced bilinguals.

2.1. Methods

2.1.1. Participants

Participants were recruited through Prolific using the following screening: Bilingual (native language + one other language, native language + two other languages, native language + three or more other languages); First language = Italian; Fluent languages = English; Language-related disorders = none; Hearing difficulties = no; Dyslexia = no; Primary language = Italian.

Among those who completed the task (N = 50), we had to remove one participant since no data were recorded during the experimental session.

To further check for L2 competence, all participants performed the LexTALE (Lemhöfer & Broersma, Reference Lemhöfer and Broersma2012), a test of lexical knowledge for English. We selected participants with a LexTALE score corresponding to a level of at least B2 (CEFR), that is, with a score (averaged % correct, % correct_av) of at least 60%. Two participants were removed from the database due to a low LexTALE score. Mean LexTALE score of the remaining participants was 82.9% (SD = 10.1%). Individual LexTALE scores are plotted in Appendix 1.A, stored on OSF.

The final sample of participants was composed of 47 bilinguals (L1 = Italian, L2 = English; 21 females, 26 males) with a mean age of 32 years (SD = 9).

2.1.2. Materials

To select the material, we started with a list of nouns in Italian corresponding either to an edible or to a not-edible object. Each noun in Italian was paired with the corresponding English noun. For each pair, we used the cross-linguistic overlap scale for phonology (Kohnert et al. Reference Kohnert, Windsor and Miller2004) to exclude cognates. We then selected 120 Italian–English pairs, 60 corresponding to an edible object and 60 to a non-edible object. An Italian L1 female speaker was videorecorded while uttering all 120 nouns in Italian.

Videos were cut so that speech began approximately six frames (i.e., 360 ms) after the video start. Videos were then modified using OpenCV in Python. First, in each video, the mouth was blurred. Then, the written English word corresponding to the audio was added, written in white on a black rectangle, and placed centred at the bottom of the video (Figure 1, top). The written word was synchronized with the speech onset. With these two manipulations, we obtained the bimodal bilingual condition (spoken Italian + written English). Then, to obtain the only speech condition, we used the same script, selecting black as the colour of the English word, so that the written word was not visible. This way, a black rectangle also appeared in the speech-only condition, and not only in the conditions with subtitles (Figure 1, bottom). To obtain the written-only condition, we muted the bimodal condition videos.

Figure 1. Top: Frame of an experimental item in the written-only or in the bilingual condition (depending on the presence of the audio) showing the English text corresponding to the Italian audio “aceto.” Bottom: Frame of an experimental item in the audio-only condition. Videos are stored on the OSF repository linked to the present study.

The 120 target items were divided into three lists of 40 items each: 20 edible and 20 non-edible. The three lists were balanced for frequency, based on film subtitles corpora (for English: Subtlex-UK, van Heuven et al. Reference van Heuven, Mandera, Keuleers and Brysbaert2014; for Italian: Subtlex-IT, Crepaldi et al., Reference Crepaldi, Amenta, Mandera, Keuleers and Brysbaert2015). For each condition, each participant received one list, with lists counterbalanced across participants, so that all items appeared in all conditions and no participant received the same item twice.

2.1.3. Procedure

The experimental task was implemented on Labvanced (Finger et al. Reference Finger, Goeke, Diekamp, Standvoss and Konig2017) and distributed online through Prolific. When entering the task, participants had to agree on data treatment. Then, they completed the LexTALE first and then the experimental tasks. For each item, they were asked to indicate whether it was edible or not edible by pressing the A and L keys.

2.1.4. Data analysis

We analysed accuracy on all items and RTs on correct answers using R version 4.3.1 (R Core team, 2021).

Before data analysis, we checked the mean accuracy by participants to exclude any participant with an overall mean accuracy lower than 80%. No participant was removed following this criterion.

Accuracy data were analysed with generalized mixed-effects regression (family: binomial, link function: logit; Jaeger, Reference Jaeger2008) with condition as fixed factor (treatment coding with bilingual bimodal [Spoken L1 + Written L2] as the reference level, to compare accuracy in the Spoken L1 condition with accuracy in the bimodal bilingual condition and accuracy in the Written L2 condition with accuracy in the bimodal bilingual condition).

RTs that were higher than 3000 ms and lower than 200 ms were excluded from the analysis. RT data were analysed with generalized mixed-effects regression (family: inverse Gaussian, link function: identity, Lo & Andrews, Reference Lo and Andrews2015). Condition was entered in the models as a fixed factor, employing treatment coding as in the accuracy analysis.

Both in the accuracy and in the RTs data analyses, we used the maximal random effect structure justified for the design, that is, random intercepts for participants and items and by participants and by items random slopes for the effect of condition ([condition|participant] + [condition|item]). If the model with this random structure did not converge, we simplified the random structure, omitting random slopes until we reached convergence, and we always avoided random structures resulting in a singular fit.

2.2. Results

Results of Experiment 1 are reported in Figure 2.

Figure 2. Boxplots showing participants’ mean accuracy and RTS by modality in Experiment 1.

The analysis of accuracy showed a significant effect of condition (χ2(2) = 28.22, p < .0001). The model’s parameters are reported in Table 1. Accuracy was lower in the Written L2 condition compared to the Spoken L1 + Written L2 condition, whereas the difference between Spoken L1 and Spoken L1 + Written L2 conditions was not significant. The analysis of RTs also showed a significant effect of condition (χ2(2) = 1905.5, p < .0001). RTs in the Spoken L1 condition were slower than in the Spoken L1 + Written L2 condition, whereas RTs in the Written L2 condition were faster than RTs in the Spoken L1 + Written L2 condition.

Table 1. Parameters of the models for Experiment 1

3. Experiment 2: Italian (L2) stimuli with English (L1) subtitles and blurred mouth

This section presents our second experiment, which has the same materials as Experiment 1, but targets a different population.

3.2. Results

Results of Experiment 2 are depicted in Figure 3.

Figure 3. Boxplots showing participants’ mean accuracy and RTS by modality in Experiment 2.

The effect of condition was significant both in the analysis of accuracy (χ2(2) = 51.40, p < .0001) and in the RT analysis (χ2(2) = 154.1, p < .0001). The models’ parameters are reported in Table 2: accuracy was lower in the Spoken L2 condition compared to the bimodal bilingual condition (Written L1 + Spoken L2), whereas there was no significant difference between accuracy in the Written L1 condition and the bimodal condition. As for RTs, responses to the bimodal bilingual condition were faster compared to the Spoken L2 condition and slower compared to the Written L1 condition.

Table 2. Parameters of the models for Experiment 2

4. Experiment 3: Italian (L1) stimuli with Italian (L1) subtitles and blurred mouth

This section presents our third experiment, which has the same population as Experiment 1, presented with the same items as in Experiment 1, but with captions in Italian instead of English.

4.2. Results

Results of Experiment 3 are shown in Figure 4.

Figure 4. Boxplots showing participants’ mean accuracy and RTS by modality in Experiment 3.

The analysis of accuracy showed no effect of condition (χ2(2) = 0.51, p < .78). The analysis of RTs, on the contrary, showed a significant effect of condition (χ2(2) = 3084.8, p < .0001). The models’ parameters are reported in Table 3: considering accuracy, no significant differences among conditions emerged. As for RTs, the bimodal condition (Spoken L1 and Written L1) was processed faster than the Spoken L1 condition, and slower than the Written L1 condition.

Table 3. Parameters of the models for Experiment 3

5. Experiment 4: Italian (L1) stimuli with English (L2) subtitles without blurred mouth

This section presents Experiment 4, which replicates Experiment 1, but with stimuli without a blurred mouth.

5.2. Results

Results of Experiment 4 are reported in Figure 5.

Figure 5. Boxplots showing participants’ mean accuracy and RTS by modality in Experiment 4.

The analyses of accuracy and RTs revealed that the effect of condition was significant (χ2(2) = 15.99, p = .0003 and χ2(2) = 1300, p < .0001, respectively). The models’ parameters are reported in Table 4: accuracy was lower in the Written L2 condition compared to the Spoken L1 + Written L2 condition. On the contrary, the difference between Spoken L1 and Spoken L1+ Written L2 conditions was not significant. Considering RTs, responses in the Spoken L1 condition were slower than in the Spoken L1 + Written L2 condition, whereas those in the Written L2 condition were faster than those in the Spoken L1 + Written L2 condition.

Table 4. Parameters of the models for Experiment 4

6. General discussion and conclusions

This study presents a series of four experiments we conducted to investigate whether processing bimodal lexical items (spoken and written stimuli) results in a cost in lexical access for unimodal bilinguals, or not. The starting point for this investigation is some recent work investigating bimodal bilinguals, which indicates that for this population, bimodal bilingual lexical access can even be faster than monolingual lexical access. Our general purpose was to shed light on the mechanisms underlying lexical access in general, beyond the bimodal bilingual population.

As Table 5 summarizes, regardless of the status of the language (whether the spoken language is L1 or L2, and the written language L1 or L2) and regardless of the bilingual status of the bimodal condition (i.e., we found the same results when the bimodal stimulus was also bilingual, as in Experiments 1, 2 and 4 and when it was not, as in Experiment 3), answers to bimodal stimuli were faster than speech-only stimuli, and slower than written-only stimuli. This stable resultFootnote ¹ suggests that: i) bimodal stimuli are processed differently from unimodal stimuli, ii) both modalities are taken into account simultaneously and iii) the integration between modalities occurs automatically, even if focusing on one modality only would have been in principle advantageous.

Table 5. Summary of the experiments and their results

As already mentioned, the rationale for this investigation was that, in bimodal bilinguals, double lexical access (spoken language + sign language) is never costly and can rather be advantageous compared to single lexical access. At first sight, we cannot say the same about double lexical access (spoken language + written language) in unimodal bilinguals, as the advantage we observe only holds with respect to the speech-only condition. A close comparison between the two sets of results deserves more discussion.

The present results, as well as the ones reported in the bimodal bilingualism literature (Donati et al., Reference Donati, Geraci, Giustolisi, Multineddu and Branchini2025; Emmorey et al., Reference Emmorey, Petrich and Gollan2012; Giustolisi et al., Reference Giustolisi, Jaber, Branchini, Geraci and Donati2024), show that bimodal stimuli are bimodally processed, that is, both modalities are taken into account, and that this bimodal process is automatic, that is, it does not depend on language dominance: it is not the case that it is activated only when it is a priori advantageous. However, the bimodality advantage is not observed in the spoken + written stimuli with respect to the written modality. What the present set of studies and those with bimodal bilinguals have in common is the presence of the spoken modality. The visual modality was, however, different, and we believe the reason why results do not converge has to do with this difference in the visual stimuli employed. Crucial points are sequentiality versus simultaneity, timing and strict or partial bimodality.

The first important factor is sequentiality versus simultaneity. Written stimuli are shown in their entirety in a single moment, whereas signed stimuli unfold over time. More generally, processing speech and processing signs is incremental due to the nature of the input, which, although with a different pace and involving different perceptual systems, is sequential. As for reading, on the other hand, visual words up to a length of 10 letters are known to be processed simultaneously (e.g., Adelman et al., Reference Adelman, Marquis and Sabatos-DeVito2010; Radeau et al., Reference Radeau, Morais, Mousty and Bertelson1992). It is thus possible that when it comes to lexical access, the reading modality introduces an advantage that can only be slowed by any simultaneous incremental processing. Some results comparing auditory, visual and audio–visual inputs in lexical decision tasks confirm this scenario (Lopez Zunini et al., Reference Lopez Zunini, Baart, Samuel and Armstrong2020)Footnote ².

A second important difference between our experiments and the ones assessing code-blending in bimodal bilinguals concerns timing. In the experiments we are discussing, written stimuli were locked with the beginning of speech, that is, items in both modalities shared the same starting point. As for sign language stimuli as well, the beginning of the sign was, in principle, synchronized with the beginning of speech (this was done spontaneously by the native bimodal bilinguals who produced the experimental stimuli). However, it is difficult to define exactly what counts as the beginning of a sign. Conventionally and from a phonological point of view, a sign begins when the hand(s) reach the place of articulation and begin(s) to move. However, some phonetic features of the sign are visible before this moment (for a discussion, see Jantunen, Reference Jantunen2015). This is well reflected by the fact that when the transition movement of the hand(s) from a relaxed position to the place of articulation associated with a sign is considered, sign identification often occurs very early and actually well before the moment corresponding to its conventional beginning (see Cecchetto et al., Reference Cecchetto, Giustolisi and Mantovan2016; Emmorey & Corina, Reference Emmorey and Corina1990). If we consider that bilinguals co-activate their languages (e.g., Thierry & Wu, Reference Thierry and Wu2007) and that this occurs also across modality (e.g., Morford et al., Reference Morford, Wilkinson, Villwock, Piñar and Kroll2011; Villameriel et al., Reference Villameriel, Costello, Giezen and Carreiras2022), the bimodal bilingual facilitation effect that is observed in lexical access with sign language–spoken language code-blendings could reflect a boost in the integration of the two modalities due to spoken language pre-activation triggered by sign language. Such an effect has no counterpart when spoken language with a written caption is considered, since no written information appears before the beginning of the audio stimulus.

A third important difference between our experiments and the ones assessing code-blending concerns the extent of bimodalityFootnote ³. While with bimodal bilinguals the two simultaneous inputs are completely independent (one being acoustic, the other visual), hence strictly bimodal, the visual items in our experiments, being words written in an alphabetic system, activate a phonological representation potentially competing with the one given with speech. The consequence is that in this setting, the bimodality is only partial. There is indeed ample evidence that healthy, skilled silent readers are unable to avoid phonological activation, even in cases in which the use of phonological information complicates the meaning resolution process (e.g., Folk, Reference Folk1999; Folk & Morris, Reference Folk and Morris1995; Van Orden, Reference Van Orden1987).

This effect of the potentially competing phonological activation prompted by written words in the bimodal condition is clearly visible if we compare the strength of our results in the bilingual Experiment 1 (see Figure 2) and in the monolingual Experiment 3 (see Figure 3)Footnote ⁴: the facilitating effect of bimodality over speech only is bigger in Experiment 3 than Experiment 1, whereas the inhibiting effect of bimodality over written only is reduced in Experiment 3 compared to Experiment 1. Remember that all the items in the bilingual experiments were controlled for their phonological resemblance, and cognates were excluded. The fact that we did find a bimodal advantage over speech, notwithstanding the “interference” of a diverging phonological activation prompted by the written stimulus, makes our result all the more significant. Not surprisingly, the advantage is amplified in the monolingual experiment (Experiment 3), where indeed the phonological activation provides a possible boosting redundancy of the stimulus.

All in all, we showed that bilinguals presented with a semantically redundant (partially) bimodal lexical item are sensitive to both modalities and process them both. This results in an advantage over speech only, which appears to be independent from language dominance, but whose size varies according to whether the setting is bilingual or monolingual.

This result encourages us to pursue exploiting the bimodal setting provided by subtitled videos to run further experiments in a number of directions. One direction is to investigate other languages, including in particular languages that do not have (fully) alphabetical writing systems, and where thus written words might prompt a reduced phonological activation, such as Korean or Chinese.

Some results discussed in Donati et al. (Reference Donati, Geraci, Giustolisi, Multineddu and Branchini2025), which did not find a consistent bimodal advantage in L2 learners of a sign language (LIS), might suggest that the factor of age of exposure might play a role in determining it. In Donati et al. (Reference Donati, Geraci, Giustolisi, Multineddu and Branchini2025), the advantage of code-blending with respect to spoken language that was identified as the crucial hallmark of native bimodal bilinguals in previous studies did not emerge in L2 learners of LIS. This difference between native and L2 hearing signers is prima facie interesting and might suggest that the advantage of lexical code-blending is out of reach for some L2 users. However, some limitations in the methodology of the study (shallow measures of frequency of LIS use and no measure of its competence) make it difficult to interpret these negative results beyond speculation.

The results presented here, which suggest that bimodal stimuli are bimodally processed in L2 learners independently from language dominance, appear to point towards a different direction and encourage us to better explore the role of age of acquisition and/or language dominance in the processing of bimodal stimuli in bimodal bilinguals.

A final direction is to replicate our experiments with stimuli that go beyond single words, manipulating the temporal alignment between words and speech (including languages sharing the same word order, as Italian and French, or different word orders, as, say, French and Korean), similarly to what has been done in Giustolisi et al. (Reference Giustolisi, Jaber, Branchini, Geraci and Donati2024) with bimodal bilinguals.

More generally, we think subtitled videos might be an important methodology to shed light on simultaneous words and sentences processing in bilinguals’ grammars.