Communication is a process in which both the speaker and the addressee express information by language, and its purpose is to convey information. Speech accommodation theory (SCT) is a sociolinguistic theory developed recently (Yuan, 1992; Ma, 1998). The theory suggests that the addressee’s speech acts characteristics can be a reference standard for the speaker to regulate speech acts. Speech convergence is a phenomenon that in daily conversation, the speaker’s speech pattern (diction, speech rate, grammar, phonology, et al.) is influenced by the addressee’s speech pattern and adjusted to it to gain the addressee’s approval and affirmation (Beebe and Giles, 1984).

Speech convergence is more likely to occur when the speaker is a subordinate or junior, and the conversation is serious. In this situation, the speaker’s speech pattern will converge with those of the addressee (Brennan and Clark, 1996; Barr and Keysar, 2002; Gijssels et al., 2016). Kemper’s study found that when young people converse with the elderly, they will slow their speech rate and reduce the length and complexity of the discourse to help the elderly understand the message, i.e., speech regulation mechanisms toward the elderly (Kemper, 1994). A study by Jiang (2017) noted that young people adopt speech convergence in telephone conversations with aged people, including but not limited to speech rate, average sentence length, and pause duration. The aim is to assist aged people in understanding the message, gain their trust, and enhance emotional intimacy.

Human-computer voice interaction design aims to simulate human-to-human verbal communication (Murad et al., 2019). Many intelligent voice devices define the voice interaction system as the user’s “intelligent voice assistant” (Rakotomalala et al., 2021) and refer to the user as the “master” during the conversation. This design reflects the relationship of ownership, subordination, and domination between the user and the voice device, comparable to the relationship between superiors and subordinates in interpersonal interactions (Nass et al., 1994; Powers and Kiesler, 2006; Ostrowski et al., 2022). On the other hand, voice interaction devices are inanimate objects, and users have lower intimacy and trust during initial use. They also interact more cautiously, which is not conducive to conveying information effectively (Dautenhahn, 2004; Höflich and El Bayed, 2015; Song et al., 2022). Thus, it is reasonable and essential to apply speech convergence strategies to improve the dialog relationship between voice interaction devices and users.

Based on speech accommodation theory, when elderly people interact with the system through voice, the system adjusts speech acts characteristics of the feedback following the elderly’s, which will enhance the emotional experience of the elderly (Myers et al., 2018, 2019). The paper analyzes the correlation between the speech rate of the elderly and their expected feedback speech rate. It also summarizes the feedback speech rate regulation strategies of the voice interaction system toward elderly people based on experimental results.

In existing research, the types of dialog tasks were divided into non-task-oriented dialogs and task-oriented dialogs according to the different purposes (Luna-Garcia et al., 2018) for which users use voice interaction devices (Chen et al., 2017).

Non-task-oriented dialogs primarily refer to forms of interaction in which users have no clear expectations or specific goals about the feedback from the voice interaction system. Typical applications include listening to music, stories, and operas. The device recognizes the type of content users want to listen to after activating the device with the wake-up word. Then the device starts to play the audio, and users enter the listening stage. Users do not initiate the voice interaction process again until the audio finishes or the audio is dissatisfying. In this usage scenario, the user’s purpose is to pass the time and relieve loneliness. When interacting with devices, users can identify only some of the information intentionally or memorize it. Users only must confirm that the device is playing the needed content, which requires a low cognitive load.

Task-oriented dialogs mainly refer to voice interaction systems that assist users in accomplishing specific tasks, such as checking the weather, making hotel or restaurant reservations, et al., by single or multiple rounds of dialogs. With a specific goal, users activate the device and give the corresponding voice command when utilizing such functions. Then users extract the information needed from the feedback played by the device and store it in short-term or long-term memory, which can be applied to the specific task. A more complex situation is that the voice interaction system conveys information to users through multi-round dialogs, such as multi-round quizzes and quiz games. In these dialogs, users must mobilize brain functions such as thinking and memory to participate in the interaction. They must also comprehend and analyze the received feedback to respond with the highest cognitive load (Sayago et al., 2019; Moore and Urakami, 2022).

As shown in Table 1, this paper conducts a voice interaction simulation experiment based on the two kinds of dialog tasks to analyze the influence of dialog task types on the expected feedback speech rate for the elderly. However, people interact with VUI with no clear goals, they also execute a dialog task. It is a little bit confusing to name this kind of interaction as a non-task-oriented dialog. So, goal-oriented, and non-goal oriented are used to describe the interactions that with and without a clear goal, respectively.

Human brain nerve cells gradually diminish after the age of 50 (Svennerholm et al., 1997; Scahill et al., 2003). Although the brain’s basic functions can generally be maintained, the volume of brain tissue may atrophy to varying degrees in some individuals, resulting in memory loss and personality changes. Brain nerve cells also impact the regulatory role of the brain’s involvement in other organs, affecting other organs’ functional performance (Wang, 2002). Compared to young people, the perceptual abilities of the elderly become blunted, and degenerative changes may occur in vision, visual perception, hearing, and auditory perception (Hawthorn, 2000; Yuan et al., 2006; Wilkinson and Cornish, 2018).

Elderly people find it more challenging to accept new things (Ziman and Walsh, 2018; Kowalski et al., 2020) and technologies (Kalimullah and Sushmitha, 2017) and acquire external information due to the deterioration in their perceptual abilities. Therefore, voice interaction simulation experiments were designed based on the feedback of different word counts from the voice interaction system to investigate the impact of the word count in single feedback on the elderly’s expected feedback speech rate.

Voice activated human machine interaction has developed rapidly in recent years. This has attracted extensive attention from the academic community. VUIs offer elderly people multiple advantages over traditional GUI/hardware interfaces by being fewer motor skills required, efficient of getting information, intuitive to interact, and rich of meaning through tone, volume, intonation, and speed (Sayago et al., 2019). More researchers are involved in the research of VUI for elderly people.

Ziman and Walsh (2018) studied the factors affecting seniors’ perceptions of VUI, and indicated that familiarity, usability, habit, aversion to typing, and efficiency of voice input are the most critical factors that influenced the seniors’ perceptions and acceptance of VUI. Research (Pradhan et al., 2020; Song et al., 2022) focused on the adoption and usage of VUI by elderly people with low technology use found that perceived usefulness, perceived ease of use, and trust are decisive factors that determined elderly people adapt VUI or not and influenced their attitudes to VUI. After studied the patterns of tactics that people employ to overcome the problems when interacting with VUI, Myers (Myers et al., 2018) indicated that feedback strategy could be a worthy point to improve VUI’s user experience.

Meena et al. (2014) proposed a data-driven approach to building models for online detection of suitable feedback response locations in the user’s speech. The results from the user evaluation through human computer interaction show that the model trained on speaker behavioral cues offers both smoother turn-transitions and more responsive system behavior. Other research that concentrated on feedback position during conversation identified feedback locations through multimodal models (Boudin et al., 2021), Interdisciplinary Corpus (Boudin, 2022), and voice activity (Truong et al., 2010; Ekstedt and Skantze, 2022).

This study applies Wizard of Oz testing to conduct voice interaction simulation experiments (Cordasco et al., 2014; Ma et al., 2022). Wizard of Oz testing is a method in which the tester acts as a “wizard” to manipulate the object to be tested to make it interact with the subject and collects relevant experimental data (Dahlbäck et al., 1993). This method is widely applied to study the usability and user acceptance of voice interaction systems, natural language applications, command languages, imaging systems, and pervasive computing applications in the prototype stage (Shin et al., 2019). In the development of voice-interactive product design, Wizard of Oz can assist UX researchers in cutting costs by quickly testing the usability of products at different stages (White and Lutter, 2003). Some studies used the Wizard of Oz testing to investigate the elderly’s acceptance of smart home products equipped with voice-interactive systems (Portet et al., 2013; Porcheron et al., 2020).

Before the experiment, dialog tasks between users and voice interaction devices were designed based on the current mainstream voice interaction process. During the experiment, subjects are first informed about the experiment’s background and the task requirements. Then subjects perform the appropriate voice interaction behavior in line with the task requirements, and the experimenter manipulates the product to provide feedback. The assistant collected subjects’ speech rate data during the process and the subjects’ expected feedback speech rate with the speech rate satisfaction scale (Ghosh et al., 2018). The assistant also needs to observe and record subjects’ performance and attitudes. Finally, user interviews and quantitative analysis were conducted to explore the users’ expectations of the feedback speech rate under different task scenarios (Iniguez et al., 2021).

To select the qualified subjects, the experimenter explained the process and conducted a pre-talk test to ensure that the subjects could hear the conversation clearly in the experimental environment and understand the overall process before the experiment. Thirty elderly people with normal hearing and no obvious cognitive impairment were chosen for the study; 10 were between 60 and 70 years old, six were 70 years old or above, and the others were 50 and 60.

The experiment was conducted in a quiet indoor environment with soft light (Figure 1). The experimenter had previously explained the experimental task process to the subjects before the experiment. The subjects sat facing the 15.6-inch display, a Philips SPA510 dual-channel speaker was used to play the voice interaction content, and a Newman ZM02 microphone collected the user’s voice commands.

The experiment was designed based on Wizard of Oz testing, in which the subject initiated the voice interaction, and the experimenter acted as a “wizard” to operate the prototype of the voice interaction system to give feedback. The prototype was made of an avatar and synthesized feedback corpus. To guide subjects, we made some slides incorporated with the prototype (see Figure 3). The dialog between subjects and voice robot prototype was simulated by switching the slides.

Take Task 1, “listen to the news,” as an example. The experimenter explained the experimental process to the subject before it started. To ensure the experiment ran smoothly, the subjects were informed of the simulated scenario of the dialog task and activated the voice interaction system in the pre-test. Slide (1) is the experiment guide, showing the subjects the dialog scenario and wake-up words. After issuing the wake-up word, subjects entered slide (2), and the experimenter played feedback, “Hey, I’m here!” indicating that the system was activated. The slide (3), an experimental guide, was shown to suggest to the subjects the dialog scenarios and voice commands to be issued. During the experiment, to enhance the subjects’ immersion, waiting and speaking animations were added to the icons of the robot.

The experiment was conducted with a 5-point Likert scale to evaluate subjects’ satisfaction with the feedback speech rate of the system (Feng, 2002). Subjects’ satisfaction was calculated based on the options corresponding to the ratings in Table 4. A higher satisfaction score of the feedback speech rate indicates a higher acceptance by the user.

While performing the experimental task, the subjects’ speech content was recorded with Adobe Audition CC2020 and a microphone in 16 kHz, mono, 32-bit audio format.

When subjects finished all dialog task, some open-ended questions were asked. The mainly purpose of this post-experiment interview is to know the attitudes and using problems of elderly people about VUI and voice robot. The questions include: “what do you think about the experience of talking with VUI or voice robot?,” “Did you have any problems when you talking with VUI or voice robot?,” “Did you always understand what the VUI or voice robot said? If not, what makes it incomprehensible?”

The number of words per second during the subjects’ speech was defined as the subject’s speech rate, noted as V_s in words/s (Li et al., 2019). Subjects’ raw recordings were imported into Adobe Audition 2020 for further processing. According to the speech rate test methodology proposed by Kim et al. (2015), subjects’ single complete speech was intercepted from the recording for speech rate analysis. If the subject paused for longer than 2 s in a single speech, the pause was deleted to obtain the corpus in the stable state, which was used to extract the subject’s speech rate. If the pause time was between 1 and 2 s, it was necessary to determine whether the subjects showed obvious doubt or nervousness according to the recorded video and choose the corpus that the subject was in a stable situation. If the subject paused for less than 1 s, it was considered normal.

The number of words is denoted as S and the total duration after deleting the silent pauses is denoted as T in single speech. The subjects’ speech rate V_s during the voice interaction is calculated by Equation (1).

Before the experiment started, speeches subjects spoke to the experimenter were recorded. Six sentences were extracted to analyze the speech rates at which subjects speak to the experimenter. At the end of the experiment, six sets of speech rate data were collected from each subject separately. The mathematical expectation speech rates at which subjects speak to a person and a voice robot are shown in Figure 4. One subject did not complete the experiment, and then there were 29 subjects’ data. A paired t-test was performed on the two kinds of speech rates, and the results show a significant difference (p = 0.000), meaning subjects speak to a voice robot at a very different speech rate than a person.

Experiments were conducted randomly with subjects on the three dialog tasks with six speech rates for feedback from the voice robot. After each session, subjects were asked to score the feedback speech rate on a satisfaction scale shown in Table 4. At the end of the experiment, the speech rate with the highest score was determined as the subjects expected feedback speech rate for this task. If more than one feedback speech rate configuration that got the highest score, then the mean of these configurations is defined as the expected feedback speech rate.

The subjects’ expected feedback speech rates of the three dialog tasks are, respectively, noted as W₁, W₂, and W₃, which are shown in Figure 5. Paired t-test was performed on the speech rate at which subjects speak to a voice robot and the expected feedback speech rate. The results showed (Table 5) that there were significant differences (p = 0.00) between the two variables for all three dialog tasks. That means in all cases of dialog task type and feedback word count, subjects expected feedback speech rate lower than their speech rate at which they speak to a voice robot.

Pearson correlation analysis was conducted on the subjects’ speech rate V_s and the subject’s expected system feedback speech rate W₁, W₂, W₃, and the results are shown in Table 6. For Task 1, the results indicate no significant correlation (r = 0.113, p = 0.56) between the subjects’ speech rate and the expected feedback speech rate. For Task 2 and Task 3, there are significant positive correlations (r = 0.417, p = 0.025 and r = 0.399, p = 0.032) between the subjects’ speech rate and the expected feedback speech rate.

The variations of the expected system feedback speech rate W₂ and W₃ with the subjects’ speech rate V_s are shown in Figures 6, 7 separately. A linear regression analysis was conducted to analyze the specific performance of the correlation between subjects’ speech rates and the expected feedback speech rates. The linear regression models are also shown in Figures 6, 7. The F-test (F = 5.670, p = 0.025; F = 5.122, p = 0.032) and T-test (t = 2.381, p = 0.025; t = 2.263, p = 0.032) also show a significance of the linear relationship of the regression model and the regression coefficient.

LSD-t-test was conducted on the subjects’ expected feedback speech rates W₁, W₂, and W₃. The results are shown in Table 7. The significance of the chi-square test was p = 0.866, which validates the homogeneity of the collected data. As shown in Table 8, a one-way ANOVA was carried out on the subjects’ expected feedback speech rates in different types of dialog tasks. The results indicate no significant difference (p = 0.065) among the three expected feedback speech rates across dialog task types. That means dialog task type did not significantly affect the expected feedback speech rate.

The post hoc LSD-t-test was conducted on the users’ expected system feedback speech rate W₁, W₂, and W₃, and the correlation between the data was examined in two pairs. The results are depicted in Table 9. There is no significant difference between W₁ and W₂ (p = 0.831) under the conditions of different types of dialog tasks and the same number of words of feedback, which also indicates that the dialog task type does not affect the elderly’s satisfaction with the expected feedback speech rate.

The feedback word count of Task 1 and Task 2 was 37, and the mean value of the subjects’ expected feedback rate was 3.61 words/s and 3.63 words/s, respectively. The word count of Task 3 was 18, and the mean value of the expected feedback speech rate was 3.81 words/s.

As shown in Table 9, the subjects’ expected feedback speech rates W₁ andW₃ are significantly different (p = 0.024). W₂ and W₃ differ significantly (p = 0.039) alike. One-way ANOVA was also carried out on the subjects’ expected feedback speech rates in different word counts of dialog tasks. As shown in Table 10, the results indicate a significant difference (p = 0.005) among the three expected feedback speech rates across feedback word count. That means feedback word count affects the expected feedback speech rate significantly.

It suggests that there was a difference in the expected speech rate of feedback between Task 1, “listen to the news,” and Task 3, “check the schedule.” Under conditions of different types of tasks and word counts of feedback, there is a difference in subjects’ expected feedback speech rates. Thus, it is essential to determine the factors affecting feedback speech rate. It indicates a difference between the subjects’ expected system feedback speech rate when completing the dialog task of acquiring information in Task 2 and Task 3, i.e., task scenarios with the same task type but different feedback word counts.

It demonstrates that the amount of system feedback words impacts users’ assessments of the system feedback speech rate. Subjects preferred a slower system feedback speech rate in the task scenario with more words of feedback than the feedback with fewer words.

Data analysis and post-experiment user interviews found that when the system feedback contained more words, the elderly’s expected feedback speech rate was significantly slower than the scenario with fewer words of feedback. It indicates that elderly people have limited cognitive ability when processing feedback information. When receiving more feedback, they take more time to remember and store the information; therefore, elderly people expect the feedback speech rate to be slower.