“You are Apple, why are you speaking to me in Turkish?”: the role of English in voice assistant interactions

4.1 Sampling and data generation

Previous studies interested in voice assistant interactions employ different approaches for generating data. Some scholars collect automatically saved interaction logs on Alexa app (Habscheid et al. 2021) and others implement an external voice recorder specifically designed for research purposes (Porcheron et al. 2018). However, the data generated with these methods only provide access to spoken interaction for a few seconds after the wake word that activates the device. Important discursive and contextual elements thus remain invisible for the analysts (Habscheid et al. 2021). Moreover, planting a voice recorder in private homes raises ethical concerns (Hector et al. 2022). Hence, in contrast to these methods, I opted for “techno-biographic interviews” (Lee 2014: 94) allowing participants to share their life stories related to voice assistant technologies. The techno-biographic interview method is rooted in traditional narrative inquiry research (Ching and Vigdor 2005; Lee 2014). Informants are encouraged to construct a narrative about their experiences with technology, much like traditional narrative interviews where individuals recount significant life events (Linde 1987; Rosenwald and Ochberg 1992). Scholars engage participants in open-ended discussions, encouraging them to reflect on their early encounters with the technology, current use cases and its impact on their lives. The approach provides insights into how people organise and interpret their subjective experiences (Linde 1987; Rosenwald and Ochberg 1992).

The participants in this study are Turkish-speaking voice assistant users who migrated from Turkey to Germany within the last 10 years. By looking at the monolingual speech design (see previous section), I argue that the participants are not considered to be the “ideal addressee” in user experience design (Piller 2001; Portmann 2022: 2). Although they may have multilingual repertoires (see Table 2), they must select only one language to use speech technologies. Unlike the smartphone assistants Siri, Google Assistant and Bixby (Samsung), Turkish is not offered as an option for the smart speaker Alexa. We thus understand that the “end-user” (Cirucci and Pruchniewska 2021) of Alexa is clearly not imagined to be someone who speaks Turkish. Looking at the one-country one-language speech design (Table 1), the user is expected to speak the national language of the country of residence, which would be Standard German in the context of Germany. These design choices, I argue, are informed by standard language ideologies and not by the multilingual realities of migrant populations.

Since I share a similar cultural, linguistic and migration background with the participants, I consider myself an “insider”, someone who exists “within the same imagined community” (Ganga and Scott 2006). This insider status has been beneficial in the recruiting process and in “making sense of the complex intricacies of situated everyday activity” (Rampton et al. 2015: 16). Nevertheless, finding Turkish-speaking voice assistant users is less self-evident than finding Turkish-speaking members of chat groups, classrooms and similar. I had to work via WhatsApp chat groups to get in touch with some of them that were willing to cooperate. In total, I could get in touch with six participants through WhatsApp chat groups, of which I am a member. These groups primarily consist of Turkish-speaking newcomers who assist each other concerning a wide variety of topics (e.g., finding housing or visa issues).

The “new wave” migrants, a term self-adopted by these digital communities, differ from the existing Turkish guest worker diaspora in Germany, in terms of socioeconomic status, age, education levels, lifestyle practices and political stance (cf. Savas 2019; Türkmen 2019). The formation of social media communities by these groups serves as a means to maintain a group identity as well-educated newcomers by differentiating themselves from the existing guest worker diaspora in Germany (cf. Oldac and Fancourt 2021). Presenting as an in-group member, I announced my research topic by asking if anyone uses voice assistants. While it was expected that the members would have higher educational levels due to the nature of these online communities, people who contacted me also work in or adjacent to the tech industry (see Table 2). This leads to believe that at least some of my informants have a rather positive view of voice assistant technology or might not be very critically predisposed to it.

The interviews were conducted through video calls in April 2021 because of the Covid-19 pandemic. I opted for a semi-structured interview with open questions (Roulston and Choi 2018), as the present study is interested in people’s beliefs and perspectives about language. In this context, it is essential to “listen to the valuations of the speakers” and orient oneself towards the interviewees by taking their individual biographies into account to avoid producing a “view from nowhere” or assuming that “linguist knows best” (Gal 2016: 131). Hence, I included questions about participants’ language repertoires and their personal history with voice assistants (see Table 2). Some of the more specific questions addressed their motivation for purchasing and using an AI assistant, their language preferences, communication issues, the gender of their voice assistants, and their personal connection to the devices. The participants’ names are withheld to protect their privacy. Table 2 assigns them fictional names, to be used when quoting interviewees during the analysis.

4.2 Coding and analytical procedure

The recorded interviews were transcribed in standard written Turkish. Paralinguistic elements such as short pauses “(.)”, laughter “@” and elongated vowels “::” are also transcribed. As this study is interested in individuals’ language ideologies, the reflexive statements of speakers – i.e., “metacommunication, participants’ talk about talk, or their reflections, signals and presuppositions about linguistic forms and their use” (Gal 2016: 116) – were coded inductively.

I applied the Grounded Theory Method for coding and categorizing to generate theory working closely with the data (Glaser and Strauss 2009). I created codes by “naming segments of data with a label that simultaneously categorizes, summarizes, and accounts for each piece of data” (Charmaz 2006: 43). These are found as subsections in the upcoming Sections 5 and 6. Whereas codes are closer to the data, describing a topic or quoting the data directly, categories are more complex (Berg and Milmeister 2011: 308). They link the codes and display connections to meta-level concepts. In what follows, I connect the instances where participants report communication issues with languages other than English (code 1) and declare English as “the original” (code 2). I show how these codes are related to technological dependencies on English in human-machine interactions (category).

Ideological constructions within these categories are examined through three semiotic processes: iconization, fractal recursivity and erasure (as discussed in Gal and Irvine 2019; Irvine and Gal 2000; Woolard 2020), which are reflected throughout the analysis. These ideological processes are realized by differentiating and comparing signs which are organized along an “axis of differentiation” (Gal and Irvine 2019: 19). In the iconization/rhematization processes, distinctions perceived in the signs are interpreted as if they mirror specific qualities and the fundamental character of speakers (Gal and Irvine 2019: 19; Woolard 2020: 10). For instance, when it comes to English-speaking voice assistants that mimic human speakers, participants associated their English with qualities such as digital, international, and natural. In juxtaposition, Turkish was linked with traits such as unnatural, robotic, a ‘follow-up’ version or perverse that were attributed to Turkish-speaking assistants. Fractal recursivity is defined as the ideological work of reproducing or repeating the contrasts created by the axis of differentiation within one of the contrasted signs (Gal and Irvine 2019: 20). Erasure, thirdly, refers to the process whereby speakers “eliminate” or “overlook” signs that refute/dispute their attempts at iconization or fractal recursivity (Woolard 2020: 10).

Considering that interviews are interactional contexts created by the researcher and the informants, it is necessary to recognize that this leads to an emergence of certain expectations regarding what interviewees should say or account for. Thus, it seems prudent to also direct attention towards the function of the language ideological work and particular associations made by participants. The data are therefore also examined with attention for how these are communicated to the interviewer, whom informants hope will understand what they are suggesting.

When it comes to the role of English in multilingual speakers’ use of voice assistants and what language ideological work speakers produce to account for their use of English in this context, we will see in the next sections that English comes with two sets of qualities. Some of these are linked to technological dependencies such as “original”, “digital” and “world language”; whereas others such as “sexy” and “natural” are associated with popular media discourses on Artificial Intelligence and standard language ideologies. In other words, one aspect is to be interpreted considering the perceived technological affordances, while the other is to be analysed in terms of its connection to science fiction narratives and the anthropomorphised design of voice assistants. As English contributes significantly to shaping local identities in the context of globalization (Lee and Jenks 2021), the analysis also explores how interviewees express their local identities through their metalinguistic reflections.

5.1 Communication issues with other languages

The participants in the study either use English or German to engage with the voice assistants (see Table 2), except for Berk who prefers Turkish. The exceptional case reveals interesting insights, as the participant emphasises that his preference for Turkish causes communication problems. In the excerpt below, he tells me about the issues that he encounters.

With his background in engineering, Berk presents himself in his account as someone who understands how the technology works or why it may cause issues with different language options, asserting that algorithms are involved in processing user inputs. As Jones (2021: 2) notes, algorithms and “our ideas about algorithms come to dominate the way we think about interacting with digital technologies”. Berk employs his understanding of algorithms to interpret the “troubles” he experiences while using voice assistants in Turkish in Germany. He evokes one language/one nation language ideologies inscribed in voice assistants by contrasting two distinct locations – a lake in Turkey “Uzungöl” and a town in Germany “Radolfzell”. These examples not only demonstrate considerable phonetic distinctions but also differences in the types of locations, highlighting the challenges posed by the technology in handling diverse linguistic and geographical contexts. What is fascinating in this case is that the “multilingual language practice” in question consists of mentioning the name of a geographical location in the territorial borders of Germany while speaking Turkish. According to his account, these circumstances led him to develop strategies to navigate the device. As Berk notes, it is easier for him to navigate Google Maps if he manually saves particular addresses, such as his workplace or home, and refers to these locations as “home” or “work”.

Five of the six interviewees (including Berk) noted difficulties communicating with their devices because they use language options other than English or speak “non-standard” English. Non-standard English in this context not only refers to deviation from “correct” pronunciation but also to using Turkish or German words. These include geographical locations, as in the above example, but also names of friends and family, or song names that the users would like to listen to (also noted in Beneteau et al. 2019).

Informants address these challenges by not only acknowledging them but also seeking ways to work around them rather than openly contest them. Berk, for instance, exonerates the voice assistant and its algorithm, choosing to navigate the difficulties by sticking to the Turkish language option and employing specific strategies for coping. In contrast, other informants lean towards English, justifying their preference as a rational choice. All participants have at least tried using their voice control devices in English, and four of them prefer it as their primary language option for voice assistants (see Table 2). In the next subsection, I explain the motivations behind their choices.

5.2 English as the “original” language

Technological innovations in voice-controlled gadgets are firstly introduced for English and, more specifically, for British and American English vernaculars. Although users are usually presented with standardized normative English variants, the voice assistants integrated into smartphones, such as Siri and Google Assistant, also offer Turkish language options. Nevertheless, the Turkish option is usually introduced after the initial release with English. For instance, Turkish-speaking Siri was released five years after the initial introduction.^[4] This fact makes ‘English as the original and standard version for voice control’ available as a resource for motivating one’s language choice. In the excerpt below, Yavuz explains his preference for English in Siri by comparing it with other language options that would come into question for him:

His reflections show that Yavuz is aware of companies’ tendency to offer more features for English, followed by other European languages like German, before providing any updates for non-Western languages such as Turkish. In the beginning of his sentence, he mentions the term “original” and then pauses shortly to specify that he is referring to English. English is notably iconized (Irvine and Gal 2000) as the original language in which users experience technologies in the most optimized way.

In the end of his statement, Yavuz explicitly differentiates Turkish from both German and English, assessing it as sounding “not natural”. As Woolard (2020: 11) notes, “iconization always trades at least implicitly in contrasts in order to generate meaning” between dichotomies that represent higher and lower scales. By iconizing English as “the original” and “world language”, Yavuz explicitly constructs Turkish as “not natural” and implicitly as merely a local language with limited widespread use. One might argue that Yavuz is not so much iconizing Turkish as unnatural but is merely alluding to the technological, robotic quality of the Turkish voice output compared to the ‘human’ voice effect that English synthetic voice use exudes. From a technical standpoint, however, all synthetic voices can be found (un)natural or (non)humanlike, depending on those who perceive them (e.g., Wagner et al. 2019). It seems more convincing, therefore, to argue that Yavuz’s perception is not so much caused by technological constraints but by his ideological stance.

Given my presence in the interview as a person who shares a similar educational background, linguistic repertoire, and migration biography with Yavuz, he also presents himself as an educated, mobile, and cosmopolitan person by highlighting that English is more “comfortable” for him as a “world” language (see also Park and Wee 2021). This presentation not only serves as an explanation of his English language preference against a researcher he thinks could question why he does not speak Turkish, which is their shared language, but it also serves as a means of indicating worldviews and values that he believes the researcher will hold, which helps to explain why he chose English.

While Yavuz explains why he prefers the language by connecting it to technological developments and his cosmopolitan identity, Talha is not sure why he favors English as the language setting of his smartphone (3). Although he prefers interacting with Siri in German, he has his phone set to English. He is used to navigating his phone visually and haptically through typed English, but not to interacting orally in English. Let us look at some of his statements from the beginning and the end of the interview:

These two excerpts from the beginning (3) and the end (4) of the interview show Talha’s thought process during our conversation. In the beginning, he reflected on his preference for English in his phone’s language settings. However, he could not give a specific reason for his choice. He reflects on the past, when he was not using a smartphone – “3310” in this context refers to Nokia’s popular phone model from the 2000s. He highlights a clear distinction between visual and oral digital engagement.

As he interacts with humans in spoken language in German and Turkish throughout the day, he finds it out of place to manage his device orally in English. Hence, he prefers interacting with Siri in German. Engaging with English-speaking voice assistants, despite not using English in daily conversations, results in a bilingual or multilingual experience that he presents as undesirable. He is alluding to discourses that emphasize that each language has its place and is to be spoken separately in appropriate contexts. A similar observation related to these ideologies was made by Heller (2002: 48) in the Canadian context, which she calls “double monolingualism”: “while bilingualism is valued, it is only valued as long as it takes the shape of ‘double-monolingualism‘. One is expected to speak each ‘language’ as though it is a homogeneous monolingual variety […]. Mixed varieties, which of course are common in bilingual settings, are frowned upon”. While Talha does not embody these ideological discourses in his personal life, he feels accountable to them and provides an explanation for his language choices. In this case, he prefers speaking one language at a time and deems it rather “strange” to “switch” his spoken language practices to English only to speak to his phone.

On the other hand, navigating the phone by reading English layout (app names, settings etc.) is not something he finds odd, although it appears to deviate from the ideological schema of “double-monolingualism” (Heller 2002). Talha’s experience reflects the insights of Gal and Irvine (2019: 21) who argue that ideologies, as “totalizing visions”, lead individuals to transform arguments that do not align with the ideological framework. Throughout the interview, Talha emphasises many “multilingual issues” or “language issues” he faces, such as the voice assistant Siri coupled with GPS navigation never seeming to work or that he gave up on using Alexa because the device could not process Turkish names, referring to one-language/one-nation technology design (as seen in Excerpt 1 by another participant). Given the fact that he frequently mentioned these issues – and that at times he did so without direct inquiry (see Excerpt 3), the informant seems to assume that it is important to address these to an interviewer who might find these insights noteworthy or relevant given her background as a Turkish-speaking university researcher in the field of linguistics. Following this narrative, he justifies the English language settings on his phone by asserting that technological affordances are to blame, as the apps are downloaded automatically in English without his interference. Later in the interview (4), he explains that his English language settings became a “habit” symbolizing his technobiography intertwined with English. Talha’s case illustrates that the standardization of English goes beyond voice control technologies. English is “naturalized” and treated as “common sense” (Fairclough 2013: 30) – a written layout that does not have to be questioned further.

Another critical aspect is which language Talha excludes in his narration. He comments on German, stating that he “never wanted it [his phone] to be German for some reason” (3). While iconizing English as the default language option, he recognizes German as an alternative option. However, he “overlooks” or “erases” (Woolard 2020) Turkish as a language option for his phone, although he explains that Turkish is a part of his language repertoire. Erasure in language ideological processes occurs when linguistic forms do not conform to the “iconic image” which speakers wish to convey (Woolard 2020: 11). By omitting Turkish and naturalizing English in this context, Talha may be presenting himself as someone closely following technological developments, often introduced in English. Thus, mentioning Turkish as an option here would not align with the kind of self-presentation he seeks. The next section addresses similar language ideologies surrounding English, showing the contribution not only of technological circumstances but also of cultural discourses about AI.

6.1 Gendered assistants

Sound in human-machine interactions seems to play an important role in the construction of computers as anthropomorphic entities. As discussed earlier in Section 2, the popular voice assistants are marketed and released with the female voice option. If the users are informed about the options and would like to change the sound, they must change the default female voice to male in the smartphone settings. However, these preferences are limited because the gender choices are only available in specific language options.^[5]

By keeping these circumstances in mind, it is not surprising that only two of the six interviewees were interested in changing the default female voice of Siri to male. The other participants were surprised by my question regarding the gendered sound for their voice assistants. The two participants, Yavuz and Sena, explain that they prefer the gendered male voice, because it appeals to them more. However, it is not only the gendered voice that leads to anthropomorphizing. They particularly find the male British English voice appealing.

Above, Yavuz and Sena tell me that they changed the voice of their assistants from female to male after trying out different language options. Multiple aspects are interesting in these statements. Firstly, the fact that they decided to change the voice option according to their attraction indicates that they attribute some human characteristics to their devices. Describing the computer voice as “sexy and nice” not only indicates that the AI assistant is depicted as a person but also an attractive one. Yavuz also notes that he wanted to show his “rainbow colors” while picking out this sound to hint at his queer identity and indicate that he finds male voices appealing. In doing so, he presents himself as a person who engages with the technology in a non-stereotypical way, opting for an unconventional gender option for Siri and wishes to emphasise his sexuality. For Yavuz, Siri functions as means to portray himself in a particular light: cosmopolitan (Excerpt 2) and queer. Similarly, Sena associates a sense of sexiness with Siri’s synthetic male voice, deliberately chosen and differentiated from Alexa’s female voice, which she categorizes as “neutral”, indexing the default setting of the technology.

Simultaneously, we see that axes other than Turkish/English are active in ideological processes when evaluating synthetic voices. British English is iconised as “sexy”, “attractive” and “nice”, differentiated from all other English languages that are associated with “fun” options to experiment with. British and American English variants as Standard English, ‘Received Pronunciation’, typically evoke indexical meanings that span from prestige, sophistication and (linguistic) authority to seriousness and intelligence (e.g., Bhatt 2002; Park and Wee 2021; Preisler 1995). The ascribed enregistered qualities may thus enhance the overall perception of attractiveness. Sena further notes that Indian and Australian Englishes are unintelligible to her, a detail that may index her educational background. Her narrative implies that her linguistic proficiency is limited to understanding ‘Standard English’ which is typically acquired through formal education in academic settings as the ‘correct’ way of speaking. The contrast drawn between British as “sexy” and Indian/Australian Englishes as “fun” thus reiterates the hegemony of Standard English within this specific cultural and technological context. The metalinguistic reflections about the choice of British English for Siri and American English for Alexa functions as a way to present oneself as educated through linguistic proficiency in (Standard) English. The following section discusses the entanglement of British English, Turkish, gender options while exploring the influence of science fiction narratives.

6.2 Science fiction narratives

The interviews show that popular media, especially science fiction narratives from the United States, influence what an AI assistant is expected to sound like. Below, Yavuz mentions Star Trek to explain why he thinks English is the appropriate language for a voice assistant:

In this excerpt, Yavuz tries out Siri’s male and female voices in Turkish to evaluate the Turkish and English language options. He was not aware that there was a male voice option for Turkish and hears it for the first time in this instance. By clicking on each option, Yavuz plays each sound separately and we hear different synthetic voices.

To understand his comments in this excerpt, it is necessary to remember that he prefers to use Siri with a British English male voice and evaluates it as sounding more attractive to him (as seen in Extract 5). After trying out Siri’s male and female voices in Turkish, Yavuz makes assessments about the Turkish and English language options. The Turkish/English axis is reiterated, in which British English is regarded as fitting and “sexy” (Extract 5) and Turkish as “perverted”. He argues that because his mother tongue is Turkish, he can point out sound design and intonation issues. This statement may be linked to the limited technological advance that has been made in Turkish voice design. Nevertheless, since there are no grammatical or intonational issues in Siri’s Turkish output, it is unlikely that the participant is referring to these formal aspects. Rather, he deems it culturally inappropriate for the AI assistant to produce synthetic voice in Turkish. Being familiar with the concept of human-machine interactions from American science fiction movies, he finds it disconcerting and strange to listen to Siri in Turkish. Through the construction of a dichotomy between these languages, qualities are assessed based on “preexisting cultural images” related to popular media culture (Gal 2016: 121). The participant picks up the qualities of a British English voice produced by an AI technology and relates them to science fiction movies like Star Trek that feature human-machine conversations in English. The synthetic voice of Siri in English is taken to index science fiction narratives, considered appropriate in the context of voice assistant interactions. Within these cultural discourses, Turkish has no place and sounds “perverted”.

Interviews with other participants also show numerous references to science fiction movies, TV programs and cartoons from the United States. These include the movie “Her” and the animated science fiction cartoon “the Jetsons”. Furthermore, some interviewees talked about science fiction narratives in general during our dialogues about voice control technologies. Although I did not ask about anything related to science fiction or mention any issue regarding this aspect, the speakers brought up some examples from movies and cartoons to illustrate their interactions. Based on these observations, there seems to be the creation of a particular cultural image based on English-speaking science fiction movies, which contributes to iconization processes of English as the standard, prestigious, natural, and culturally appropriate language of digital assistants. These circumstances would also suggest that oral interactions with other languages such as Turkish may be interpreted as strange and unnatural.

Another issue that Yavuz mentions is that it seems odd for him to interact with an Apple gadget in Turkish (7), which may result from the company’s origin in the United States. Hence, the voice assistant is not only perceived as a conversational AI associated with science fiction narratives, but it also represents the company that markets the technology. When the US-based company speaks through the AI assistant in Turkish, the company’s authenticity is questioned, leading to iconization of Turkish as culturally inappropriate.

Additionally, Yavuz grapples with inconsistencies in the iconic image associated with English. While he favours the British English voice option, he provides examples tied to American culture. This incongruity prompts the reflection that he may associate British English with the broader anglophone culture, perceiving it as the standard form – akin to received pronunciation.

While certain dependencies on English in technologies are acknowledged due to the inherent technology design (discussed in the previous section), the participants leverage English-speaking voice technologies entangled with science fiction narratives and standard language ideologies to shape a particular image of themselves – a portrayal marked by a cosmopolitan, mobile, educated, and technologically savvy lifestyle. This phenomenon can be attributed to the specific data sampling and participant recruitment methods employed. As outlined in the methodology, I engaged with these individuals through digital chat groups known as the “New Wave” of migrants in Germany. These communities identify themselves as comprised of newcomers with a high level of education. The announcement of my project concerning voice assistants likely resonated with those who consider themselves as actively engaged with contemporary AI voice technologies and enthusiasts of technological advancements. Furthermore, while interacting with a researcher who shares a similar linguistic and migration background as themselves, informants are accounting for the prominent or continued role of English in their voice assistant use, by means of ideological associations that they expect the interviewer to understand or find relevant.