2.1. Method

Design and Stimuli. Experiment 1 had a 2 (deictic shift: shift versus no shift) $ \times $ 2 (speech type: direct speech versus indirect speech) within-subjects design. The dependent variable was speech latency (from the onset of the answer screen to the onset of the speech report). The independent variable deictic shift was manipulated to determine whether the deictic words in an original utterance needed to be changed when reporting indirectly. For example, imagine John wants to quote Linda’s utterance, ‘Mary had a pony when she was a little girl’. The deictic words in direct and indirect speech are the same. Therefore, no deictic shift is needed. If Linda says, ‘I had a pony when I was a little girl’, the deictic word ‘I’ needs to be changed to ‘She’ when John quotes Linda indirectly. For the experiment, 40 experimental stories were created as stimuli, among which 20 stories were for the deictic shift condition and the other 20 stories were for the no deictic shift condition. Each story consisted of four sentences in which the last two sentences were always presented as a dialogue between two people. All stories were presented in the form of a picture (Fig. 1).

Figure 1. An example of the stories from the deictic shift condition in Experiment 1.

Participants. We were interested in the interaction between deictic shift and speech type. The determination of the sample size was based on Brysbaert’s (Reference Brysbaert2019) power analysis investigation for a within-subjects design with two factors. 110 participants are needed to detect an interaction effect in a two-factor repeated-measures design with an effect size of d = 0.4 and a power of 80% (Brysbaert, Reference Brysbaert2019). We recruited 120 participants to have enough observations after removing invalid data. Six participants were excluded due to a program crash. This resulted in a final sample of 114 participants (M_age = 20.21; 79 females). All participants were either native English speakers or had a minimum level of 80 on the Test of English as a Foreign Language (TOEFL) or a 6.0 on the International English Language Testing System (IELTS). They received 1-hour research hour credit for participation. Informed consent was obtained from all participants before the experiment. This experiment has been approved by the Research Ethics Review Committee DPECS of the Erasmus University Rotterdam.

Procedure. Each participant was tested individually in a sound-attenuated room. First, participants read the instructionsFootnote ³ and completed eight practice trials to become familiar with the procedure. The instructions indicated that they would be presented with a short story and thereafter needed to answer a question that corresponded to the story. The purpose of Experiment 1 was to test the production of direct and indirect speech when the verbatim memory of a to-be-reported utterance was accessible. Therefore, participants were asked to read the story carefully and memorize the last sentence of the story. In the deictic shift condition, the last sentence always contained at least two deictic terms that needed to be transformed when participants produced indirect speech.

Each trial began with a fixation cross appearing in the middle of the screen for 500 ms, followed by the story. After inspecting the story and memorizing the last sentence, participants pressed the SPACE bar to see the question. In total, 40 experimental questions were created to elicit direct or indirect speech (e.g., What did Marie say?). The questions were presented for 2,000 ms and then replaced by the answer screen. The first part of the answer was provided to prompt direct and indirect speech (Fig. 2). The participant’s task was to complete the sentence according to the story they had just read. For example, when the question was ‘What did Marie say?’ participants would be presented with either ‘She said:’ to elicit direct speech or ‘She said that’ to elicit indirect speech. Participants were asked to recall the sentence verbatim. This means that no word from the original sentence should be changed when producing direct speech, and only deictic terms were changed in indirect speech. To make the utterance planning more extensive, participants were instructed to speak as fluently as possible. Participants answered the questions by speaking into a microphone. Speech latencies and audio responses were recorded for later analyses. Data and materials can be accessed via https://osf.io/zha4t/?view_only=e9e0b8ec491d4af1bf1c0e6c605ca7c2.

Figure 2. Experiment procedure for Experiment 1.

2.2. Statistical analysis

Several types of responses were excluded from the final data analyses. First, trials with wrong responses (i.e., trials in which participants failed to recall verbatim or made errors in transforming deictic terms) were removed (8.3%). Second, utterances with disfluencies were removed (1.8%). Applying these exclusion criteria resulted in the removal of 10.1% of the trials in total. Data analyses were conducted with R version 4.2.2 (R Core Team, 2022) and lme4 version 1.1.31 (Bates et al., Reference Bates, Mächler, Bolker and Walker2015) to perform a linear mixed-effects analysis of the influence of deictic shift and speech type on speech latencies of reported speech (Baayen et al., Reference Baayen, Davidson and Bates2008). The model contained deictic shift and speech type and their interaction as fixed effects. As for random effects, we entered the participants and stimuli as varying intercepts, as well as by-participant and by-stimulus random slopes for the effect of deictic shift. P-values were obtained through likelihood ratio tests, comparing the full model including the specific variable to a reduced model that does not include the variable.

4.1. Method

Design and Stimuli. Experiment 2 used a 2 (interference: yes versus no) $ \times $ 2 (speech type: direct speech versus indirect speech) mixed design. Interval was a between-subjects variable, with half of the participants needing to finish an interference task before the sentence completion task and the other half of participants only finishing the sentence completion task. Speech type was a within-subjects variable. A total of 20 stories were created as stimuli. Similar to Experiment 1, each story consisted of four sentences, and the last two sentences were presented as a dialogue between two people. The incomplete sentence was created by removing the first word of the last sentence in the dialogue.

Participants. According to Brysbaert (Reference Brysbaert2019), for a two-way ANOVA with one within-groups factor and one between-groups factor, 67 participants for each group are needed to run a study with 80% power. To ensure a sufficient number of participants after removing invalid data, we recruited 150 university students as participants. Five participants from the interval condition were excluded due to a program crash, and the last five participants from the no-interval condition were removed to ensure equal numbers of participants in both conditions. The final sample size consisted of 140 participants (M_age = 22.15, 92 females). All participants were either native English speakers or had a minimum level of 80 on the Test of English as a Foreign Language (TOEFL) or a 6.0 on the International English Language Testing System (IELTS). They received 1-hour research hour credit for participation. Informed consent was obtained before the experiment. This experiment has been approved by the Research Ethics Review Committee DPECS of the Erasmus University Rotterdam.

Procedure. Participants were tested individually in a sound-attenuated room. After reading the instructions and indicating that they already understood the procedures, participants were asked to finish six practice trials to be familiarized with the task. The procedures began with a fixation at the center of the screen, reminding participants that the experiment would start soon. Following the fixation, participants were presented with a short story. They were instructed to read the story carefully and press the SPACE bar after they had understood it. After reading the story, half of the participants continued to complete a counting backward task. In the counting task, a 3-digit number (e.g., 469) appeared on the screen. Participants were instructed to first read out this number and then count backward in steps of 3. Participants kept counting for 20 seconds until they saw the word ‘STOP’ on the screen. Following the stop sign, a fixation appeared again on the screen for 200 ms, and an incomplete sentence was presented. Participants were asked to think of a word that could make this sentence complete based on the story they had just read. Once they had the answer, they read the whole sentence into a microphone as fluently as possible. Participants immediately finished the sentence completion task after reading the stories for the no-intervention condition. Speech latencies and audio responses were recorded for later analyses.

4.2. Statistical analysis

Responses with wrong answers (9.87%) or disfluencies (6.92%) were removed. Data analyses were conducted with R version 4.2.2 (R Core Team, 2022) and lme4 version 1.1.31 (Bates et al., Reference Bates, Mächler, Bolker and Walker2015) to perform a linear mixed-effects analysis of the influence of interval and speech type on speech latencies of reported speech(Baayen et al., Reference Baayen, Davidson and Bates2008). The model contained interference and speech type and their interaction as fixed effects. In our analysis, we accounted for random effects by including participants and stimuli as varying intercepts. Additionally, we included by-participant and by-stimulus random slopes for the effects of speech type. P-values were obtained through likelihood ratio tests, comparing the full model including the specific variable to a reduced model that does not include the variable.

4.3. Results and discussion of Experiment 2

There was a significant interaction between interference and speech type ((χ ²(1) = 24.86, p < 0.01). A post-hoc analysis showed that, in the interval condition, speech latencies of indirect speech (M = 1619 ms, SD = 609 ms) were significantly shorter than speech latencies of direct speech (M = 1886 ms, SD = 764 ms) (p < 0.01). In the no-interval condition, speech latencies of indirect speech (M = 1137 ms, SD = 311 ms) did not significantly differ from speech latencies of direct speech (M = 1096 ms, SD = 229 ms) (p > 0.05) (Fig. 4). Our hypotheses were only partially supported by these results. In line with our prediction, when the verbatim trace of an utterance was impaired by an intervening task, recalling the utterance using indirect speech was faster than direct speech. To our surprise, Experiment 2 did not observe a difference in speech latencies for direct and indirect speech when verbatim memory was available. This result seems to be in contrast with the results from Experiment 1, where we observed that direct speech production was faster than indirect speech, but it can be explained by the differences in tasks. We will return to this discrepancy in the discussion section.

Figure 4. Log reaction time in each condition of Experiment 2.

4.4. General discussion

The current study examined whether and how memory representation and deictic shifts influenced direct and indirect speech production difficulties. We compared production difficulties of direct and indirect speech in conditions where either a deictic shift was or was not required to produce reported speech. In both experiments, participants read short dialogues between two interlocutors and were prompted to answer questions using either direct or indirect speech. Experiment 1 investigated the effect of deictic shifts on production difficulties when participants had verbatim memory of the to-be-reported utterances. The results showed that deictic shifts significantly influenced speech latencies. When indirect speech production required a deictic shift, participants took a longer time to initiate indirect speech than direct speech. When no deictic shifts were needed, direct and indirect speech latencies were similar. In Experiment 2, we continued to investigate the effect of deictic shifts on production difficulties when participants’ verbatim memory of to-be-reported utterances was interfered by an intervening task. The results showed that when verbatim memory was hindered, producing direct speech was slower than indirect speech. Unlike Experiment 1, we did not observe a difference in speech latencies between direct speech and indirect speech when verbatim memory was available.

This study was motivated by several arguments and observations, suggesting that direct speech is an easier mode of reporting than indirect speech (Groenewold et al., Reference Groenewold, Bastiaanse, Nickels, Wieling and Huiskes2014; Li, Reference Li and Coulmas1986). The results of our study, however, painted a more complicated picture. First, we argued that the memory representation of the to-be-reported utterances should play a role in the production of reported speech because reported speech involves a recollection of previously said content. Even though direct speech may not necessarily be a verbatim replication of to-be-reported content, the retention of surface structures plays an essential role in language production. Verbatim memory can be used in the immediate output (Garrod & Trabasso, Reference Garrod and Trabasso1973). Experiment 1 tentatively supported the role of memory in reported speech production, showing that direct speech production was faster than indirect speech when verbatim memory was readily accessible to participants. We predict that the reason for the longer speech latency for indirect speech is that it requires a transformation of deictic words, whereas direct speech does not. When direct and indirect speech shared the same deictic expressions (e.g., She said: ‘John was late’. versus She said that John was late.), the production of these two reporting styles was equally fast. This result again supports the argument that when verbatim memory was accessible to participants, the difference in speech latencies was caused by deictic shifts. When no deictic shift was required, there was no difference between direct speech and indirect speech in terms of speech latencies. In sum, the findings from Experiment 1 demonstrate two points: First, the retention of verbatim memory facilitates the production of direct speech. Second, verbatim retention only has a facilitating effect when the production of indirect speech requires a transformation of deictic expressions. When the deictic terms are the same in direct and indirect speech, there is no difference in speech latencies.

Experiment 1 helps to clarify the existing view arguing that direct speech is easier to produce because it is a verbatim replication of previous utterances (Li, Reference Li and Coulmas1986). While this belief has been criticized because people do not always report verbatim (Wade & Clark, Reference Wade and Clark1993), we cannot deny that, between direct speech and indirect speech, only direct speech allows people to recite in a verbatim manner. Our findings from Experiment 1 modify this view by suggesting that, when people have verbatim memory of previous utterances, direct speech is easier than indirect speech because no deictic shift is needed for direct speech production.

Previous studies repeatedly show that the retention of an utterance’s surface form decays quickly and is prone to interference (Fisher & Radvansky, Reference Fisher and Radvansky2018; Sachs, Reference Sachs1967). In most daily communication circumstances, the production of direct and indirect speech can only rely on the gist memory of previously said utterances. In Experiment 2, we compared speech latencies of direct and indirect speech when verbatim memory was not disturbed by an interfering task. We introduced a counting backward task that came in between the dialogue reading and the speech production task. We expected the production of direct speech to be slower than indirect speech when the verbatim trace was disrupted. We observed that when participants’ verbatim memory of an utterance was hindered, indirect speech production was faster than direct speech. This aligns with previous findings from reported speech comprehension studies showing that participants spent more time interpreting pronouns in direct speech than in indirect speech (Köder et al., Reference Köder, Maier and Hendriks2015; Köder & Maier, Reference Köder and Maier2016). Resolving pronouns in direct speech requires a shift from the third-person perspective to the first-person perspective, resulting in a longer resolution time (Köder et al., Reference Köder, Maier and Hendriks2015; Köder & Maier, Reference Köder and Maier2016). Similarly, in Experiment 2, the production of direct speech also required a shift from the current situation to the time and space in which the original utterance was produced. This shift leads to longer reaction times for direct speech production than for indirect speech. The finding from Experiment 2, together with prior reported speech comprehension studies, suggests that perspective shift contributes to processing difficulty differences between direct speech and indirect speech.

Unlike Experiment 1, we did not observe significant differences in speech latencies when there was verbatim memory available. This is unexpected, considering that the production of direct speech was significantly faster than indirect speech in Experiment 1 when participants had access to verbatim memory. We speculate that this could be because different tasks were used in Experiment 1 and Experiment 2. In Experiment 1, participants recalled the whole sentence. Each sentence contained at least two deictic terms that needed to be changed to produce indirect speech. In contrast, in Experiment 2, participants only needed to paraphrase one pronoun. Presumably, the transformation of only one deictic term is not very resource-consuming and therefore can happen at a fast pace. Because the task demand of Experiment 2 was less taxing than that of Experiment 1, no difference in speech latencies was detected when verbatim memory was available in Experiment 2.

We recognize several limitations of our study. Firstly, we note that deictic shifts might only account for part of the production difficulty differences between direct speech and indirect speech. This study did not test whether other syntactic differences between these two reporting styles would influence their production difficulties. Some sentence structures have been demonstrated to be more difficult to produce than others, as is reflected by longer speech latencies or a larger pupil size (Altmann & Kemper, Reference Altmann and Kemper2006; Sevilla et al., Reference Sevilla, Maldonado and Shalóm2014). For example, producing passive and object-dislocated sentences evokes larger pupil dilation than active–canonical subject–verb–object sentences, suggesting that passive–noncanonical structures are more difficult to produce and require more cognitive effort (Sevilla et al., Reference Sevilla, Maldonado and Shalóm2014). As discussed before, direct speech and indirect speech have different sentence structures, with direct speech constructed as a main clause and indirect speech as a subordinate clause. In a main clause, to-be-reported content is directly attached to the reporting verbs (e.g., say). Consequently, direct speech can convey all syntactic structures, whereas indirect speech can only convey utterances that are grammatically correct in a subordinate clause. Observational and empirical evidence showed that people tend to use direct speech more frequently when the to-be-reported utterances are perceived as grammatically less acceptable in indirect speech (Mayes, Reference Mayes1990; Li et al., Reference Li, Jongerling, Dijkstra and Zwaan2022). The use of direct speech sometimes involves much less syntactic paraphrasing than indirect speech. For instance, suppose A would like to quote B’s utterance ‘Say what?’. A can say, ‘B said, “Say what?”’ in direct speech. A can also use indirect speech and say, ‘B asked if I could repeat what I said’. One would expect that when indirect speech requires much syntactic paraphrasing, it will be more difficult to produce compared with direct speech.

Second, we did not test the influence of non-verbal information on the production of direct and indirect speech. As demonstrated by prior studies, the use of non-verbal information might facilitate speech production, especially for people with limited language competence (Groenewold et al., Reference Groenewold, Bastiaanse, Nickels, Wieling and Huiskes2014; Ulatowska et al., Reference Ulatowska, Reyes, Santos and Worle2011). The non-verbal information could facilitate the production of direct speech by either replacing a difficult word with gestures or helping to recall the content. Despite this limitation, the results of our study remain informative because people often encounter direct speech without non-verbal information involved, such as producing direct speech in a written form.

Finally, this study did not consider reported speech production difficulties in various discourse types. Prior studies show that discourse types might influence the difficulty of shifting perspectives. In particular, people make more mistakes and have longer reaction times when interpreting pronouns in direct speech than in indirect speech because comprehending direct speech requires a deictic shift. However, people’s abilities to interpret pronouns in direct speech are significantly improved when direct speech is embedded in a narrative setting compared with an information transmission setting (Köder et al., Reference Köder, Maier and Hendriks2015; Köder & Maier, Reference Köder and Maier2016). The narrative setting makes it easier to take the first-person setting and therefore reduces the time for perspective shifting (Köder & Maier, Reference Köder and Maier2016). Future research can compare whether different reporting contexts will also influence the difficulty of perspective shifting in reported speech production.

As discussed in the limitation section, the comparison of production difficulties is an extremely complicated topic. The relative production difficulty of direct and indirect speech may depend on factors such as the memory representation of the utterances, the properties of the original utterances, or the contexts in which reported speech is produced. Despite the above-mentioned limitations, this study contributes to further our knowledge and understanding of direct and indirect speech. The most important theoretical contribution is that we provided empirical evidence for the role of memory representation in the production of direct and indirect speech. We can conclude that the availability of verbatim memory facilitates the production of direct speech, reflected by shorter speech latencies compared with indirect speech. When the verbatim representation is hindered, the production of indirect speech is faster instead.