team
Solo Project
Role
UX Researcher
duration
3 Months
Tools & Methods
Literature review
sociolinguistic variation in conversational ux design
Examining the intersection of chatbots, voice interfaces, and sociolinguistics through a literature review.
overview
About the project
Linguistics (especially sociolinguistics) and UX are two fields I'm incredibly passionate about, so when I had to write a literature review for a course, I knew I wanted to work at the intersection of the two. The result synthesizes nine studies to understand where conversational interfaces, like chatbots and voice assistants, might break down for anyone who doesn't speak in a prescriptively, monolingually "standard" way, and how sociolinguistic concepts can offer solutions.
The following summarizes my work, but if you'd like to see more, you can find the full report here. Enjoy reading, and please reach out if you'd like to discuss any thoughts :) !!
One flag up front: the literature here predates the recent wave of LLM-based assistants, which, I would argue, could make the core question sharper, rather than dated. More on that below.
Context
Where conversational user experiences fall short
Conversational User Experiences (CUX) aim to create interactions that feel human. However, human speech isn't uniform; it shifts across communities, contexts, and accents. Most CUXs are tuned to a single, standard-English speaker, so they stumble with non-native speakers, multilinguals who may code-switch, and multi-person conversations. The cost is users' satisfaction, confidence, efficiency, trust, and adoption of these tools. This review explores how sociolinguistics (in short: the societal impact on language, through social interactions, dialects, accents, code-switching, and listener preferences of speech patterns) can close that gap, and briefly discusses the implications of technolingualism, or the role technology plays in cultural stereotypes.
Research
Three places sociolinguistics can help CUX design
Insight 01
Inputs: understanding the user
Design for multi-speaker awareness, broader language support, and room for code-switching
CUX struggles to process speech that isn't "standard." Voice assistants can't reliably tell who's speaking in a group, or signal what they've failed to understand when asking you to repeat yourself (Koh, 2021), which can result in each conversational participant having conflicting views of appropriate responses.
Non-native English speakers also hit walls with vocabulary, sentence structure, and pronunciation (Pyae & Scifleet, 2019), and bilingual users may code-switch, whether or not the system can follow – but complete tasks more successfully when it can (Parekh et al., 2020).
Insight 02
Outputs: how the interface speaks
Outputs benefit from individualization and context-awareness
Linguistic style, register, and accent impact perceived competence and trust (Chaves et al., 2022), but this preference isn't universally generalizable; it's shaped by colonial histories, social class structures, and individuals' experiences. Some users may favor a voice that sounds like one from their own community, while many Singaporean participants, for example, ranked British accents higher than local ones on trust and politeness (Niculescu et al., 2008).
Because preferences split in various ways, outputs can benefit from letting users shape the voice, or automatically adapting to contexts, rather than defaulting to a "neutral" setting.
Insight 03
Ethics: voice is never neutral
Deliberately diversify outputs to reduce biased perceptions
Because technology both shapes and is shaped by language – what Pfrehm (2018, as cited in Sutton et al., 2019) calls technolingualism – the accent or style a system defaults to carries cultural weight, and can quietly reinforce which ways of speaking get treated as "prestigious" and which don't (Sutton et al., 2019). To address this, they suggest using new or unfamiliar voices and/or accents, or reducing human-likeness.
Conclusion
The takeaway
“VUIs have achieved such a level of technical capability that attention can move towards considering [outputs rather than processing] in VUI design…now would be an appropriate time to explore relevant knowledge from other disciplines”
Sutton et al. (2019)
Sociolinguistics offers the most well-defined route to making conversational UX work for international and multilingual audiences — not as a final polish, but from the start; ensuring trust and effectiveness early drives adoption, so input processing (multi-user, accents, languages, code-switching) and output style (register, prestige, context) are worth improving concurrently rather than sequentially.
REFERENCES
Retrospective
Where this stands today
The systems critiqued by these studies were mostly intent-based assistants, like early Alexa, Google Home, and Siri. LLM-powered tools today are far stronger at processing and producing varied speech, so it's fair to ask whether these analyses still hold.
More than anything, the current state of LLMs particularly strengthens the technolingualism consideration: a language model inherently learns the dominant language variety of its training data (usually a narrow, standardized one), so it tends to understand under-represented speakers less accurately and can echo the stereotypes baked into web-scale text (Grieve et al., 2024). Some tools can also qualitatively analyze speech patterns and suggest improvements – these most frequently penalize Black and female speakers, while nudging others to a more "standard" sound (Holliday & Reed, 2025).
This raises some questions: whose variety is the model built on, and who might get misunderstood because of it? When a tool offers to soften or improve speech styles, whose standard are you moving toward? As more of what we say and write passes through these systems, does the range of ways we're allowed to sound quietly narrow?
Better fluency for standard speakers can widen the gap for everyone else, unless variation is accounted for as a design goal rather than an afterthought, which is what the original literature review points toward; while these sociolinguistic concepts may apply to different problems today, the questions and solutions remain the same.
