Code-Switching in Multilingual Interviews: 2026 Guide
Qualitati Research Team · 2026-08-21 · 9 min read
Short answer: Code-switching — a participant mixing two languages inside one sentence — is where multilingual research pipelines break. Transcription errors cluster at switch points, and language models handle mixed text asymmetrically. As of August 2026, the fix is not a better model but a workflow: label the language mix, audit switch-point accuracy, and keep analysis in the participant's dominant language.
Why code-switching matters in multilingual interviews
Ask a bilingual product manager in Barcelona how she uses your dashboard and you will not get clean Spanish. You will get Spanish grammar carrying English nouns: "el onboarding fue confuso, pero el dashboard sí me gustó." This is code-switching, and among bilingual speakers it is the norm, not an edge case.
For qualitative research this creates a specific, under-discussed risk. Most transcription and analysis tooling assumes one language per file. When the speaker switches, several layers fail at once — the acoustic model, the tokenizer, and the downstream coding step — and they fail precisely on the words that carry the product meaning, because those are usually the borrowed English terms.
The result is a quiet data-quality problem. The transcript looks fluent. The themes look plausible. But the sentences where participants said the most useful thing are the sentences most likely to have been garbled.
Key takeaways
- Code-switching degrades automatic speech recognition unevenly across systems — a 2026 ServiceNow-AI benchmark of seven frontier ASR models found top systems take only a small penalty while others degrade sharply, which means model choice matters more than it does for monolingual audio.
- Language models process mixed text asymmetrically: inserting foreign tokens into English hurts comprehension, while embedding English into another language often helps.
- Errors concentrate at switch points, so a corpus-level word error rate hides the problem. Measure accuracy in the two-to-three-word window around each switch.
- Loanwords are not code-switching. Conflating them inflates your error estimates and wastes review time.
- Practical mitigation is procedural: declare the expected language pair up front, audit a switch-point sample, analyze in the matrix language, and have a bilingual reviewer verify every quote you publish.
What the 2026 evidence actually shows
Speech recognition: robustness varies more than difficulty
In a benchmark published on 9 June 2026, researchers at ServiceNow-AI evaluated seven frontier speech-recognition systems — including AssemblyAI Universal 3-Pro, Deepgram Nova 3 Multilang, ElevenLabs Scribe V2, Google Gemini 3 Flash, Mistral Voxtral Small, Nvidia Parakeet TDT, and OpenAI Whisper Large V3 Turbo — on Spanish-English, French-English, Canadian French-English, and German-English code-switched speech. Each utterance was compared against monolingual matrix-language and monolingual English baselines (ServiceNow-AI, 2026).
The headline finding is not that code-switching is uniformly hard. It is that code-switching separates robust systems from fragile ones. Top performers incurred only a small penalty relative to monolingual baselines; weaker systems degraded substantially, with Whisper showing the largest relative degradation on the German-English pair. For a research team, that reframes the decision: you cannot pick a transcription vendor on a monolingual accuracy chart and assume the ranking holds for your bilingual participants.
Where the errors live: switch points and loanwords
The HiKE evaluation framework (Paik et al., submitted September 2025, revised January 2026) makes a methodological point that transfers directly to research practice. HiKE is a non-synthetic Korean-English code-switching benchmark, and its contribution is the labeling scheme: it annotates code-switching hierarchically at word, phrase, and sentence level, and it labels loanwords separately.
That separation matters. "Feedback" spoken inside a Korean sentence may be an established loanword rather than a genuine switch, and treating the two identically muddies any accuracy estimate you compute. HiKE also reports that multilingual ASR models initially underperform on code-switched material but improve with targeted fine-tuning — useful context, but not something most research teams will do themselves.
Analysis: the asymmetry nobody plans for
Transcription is only half the pipeline. In "Lost in the Mix: Evaluating LLM Understanding of Code-Switched Text" (Mohamed et al., June 2025), the authors built code-switched variants of standard reasoning and comprehension benchmarks and found a directional effect: when foreign-language tokens interrupt English text, model performance degrades noticeably, but embedding English inside another language frequently improves comprehension.
Read that as a workflow instruction. A Spanish transcript sprinkled with English product terms is the favorable direction. An English transcript with Spanish phrases dropped in is the unfavorable one. If your standard operating procedure is "translate everything to English first, then code," you may be converting the easier case into the harder one. The same paper reports that prompting-based mitigation gave mixed results while fine-tuning was more stable — which, for teams without a fine-tuning budget, means prompt tricks are not a reliable fix.
The wider picture is consistent. A survey of 327 studies across 15+ NLP tasks, 30+ datasets, and 80+ languages concludes that most large language models still struggle with mixed-language input, and that code-switching datasets and evaluation methods remain limited (Sheth et al., 2025). Conversational, speech, and multimodal code-switching are singled out as the least-developed areas — which is precisely the territory qualitative interviews occupy.
Participants switch languages on purpose
One more finding reframes code-switching as signal rather than noise. An extended abstract at the 2026 CHI Conference reports a mixed-methods study of 84 multilingual users — combining surveys, semi-structured interviews, and interaction logs — examining how they choose and switch languages when talking to LLMs. The preliminary finding is that code-switching is a strategic practice, used to manage task efficiency, linguistic precision, and emotional comfort (CHI 2026 Extended Abstracts).
If that holds for research interviews too — and interview methodologists have long argued something similar — then a switch into the participant's first language often marks the emotionally loaded moment in the transcript. Flattening it into monolingual output destroys exactly the passage you most wanted.
The Code-Switching Readiness Audit
This is a Qualitati framework for teams running interviews with bilingual participants. Work through it before fielding, not after.
| # | Step | What to do | Pass condition |
|---|
| 1 | Declare the pair | Name the expected matrix language and embedded language per segment (e.g. Spanish matrix, English embedded) | Every recruited segment has a declared pair before fielding |
| 2 | Separate loanwords | List the borrowed terms your domain uses natively in the matrix language and mark them as loanwords, not switches | A written loanword list exists per language pair |
| 3 | Pilot for robustness | Transcribe 3–5 pilot recordings on your candidate systems and compare them on the same audio | Chosen system verified on code-switched audio, not a monolingual benchmark |
| 4 | Audit switch points | Hand-check the 2–3 word window before, during, and after each switch in a random sample | Switch-point accuracy reported separately from overall accuracy |
| 5 | Choose the analysis language | Code in the matrix language where possible; avoid pre-translating into English | A documented decision, with rationale, in the study protocol |
| 6 | Verify published quotes | A bilingual reviewer checks every quote against the audio, not the transcript | 100% of published quotes verified against source audio |
Switch-point risk matrix
Use this to decide how much manual review a study needs.
| Language distance | Low switch frequency | High switch frequency |
|---|
| Close pair (e.g. Spanish-English) | Low risk — spot-check quotes only | Moderate risk — audit a switch-point sample |
| Distant pair (e.g. Chinese-English, Arabic-English) | Moderate risk — audit a switch-point sample | High risk — full bilingual review of any transcript used for quotes |
Who this is for
Research teams interviewing bilingual or multilingual participants: consumer insights in Latin America, Southeast Asia, the Gulf, or India; UX research with immigrant or diaspora users; enterprise research where domain vocabulary is English but daily conversation is not; and any team running the same study across several markets and comparing themes across them.
When not to use this approach
If your participants are genuinely monolingual, this audit adds cost without benefit — run your standard pipeline. If your study is legally or clinically sensitive, automated transcription of code-switched speech is not the right starting point at all; commission professional bilingual transcription and use AI only for downstream synthesis. And if you need a defensible published claim about switch-point accuracy in your specific language pair, treat the audit above as an internal quality gate rather than as evidence — it is a review procedure, not a validation study.
Where Qualitati fits
Qualitati is an AI user research platform for product, UX, and customer insights teams, supporting research in ten languages: English, Chinese, French, Norwegian, Dutch, German, Spanish, Portuguese, Japanese, and Arabic.
Three parts of the platform bear directly on the workflow above. AI-moderated interviews can be run in the participant's own language rather than forcing English, which keeps the matrix language intact — the favorable direction for downstream analysis. ThemeLens runs thematic analysis across up to 100 transcripts in a map-reduce pipeline, mapping codes to research questions and anchoring themes in participant quotes, so every theme keeps a traceable line back to a specific passage you can check. And the QDA Workspace supports human-in-the-loop inductive and deductive coding, which is where a bilingual reviewer does step 6 of the audit.
What Qualitati does not do is remove the need for the audit. No platform can guarantee switch-point transcription accuracy for an arbitrary language pair, and any vendor claiming otherwise is describing a monolingual benchmark. The honest position, as of August 2026, is that code-switched research needs a human verification step, and the platform's job is to make that step cheap and traceable rather than to pretend it is unnecessary.
Limitations and open questions
Three caveats belong on this article. First, the ASR benchmark discussed above covers European language pairs; its ranking should not be assumed to transfer to Chinese-English, Arabic-English, or Hindi-English. Second, the CHI finding on strategic code-switching is reported as preliminary in an extended abstract and concerns human-LLM chat, not research interviews — the parallel to interviewing is an inference, and a reasonable person could disagree with it. Third, "Lost in the Mix" evaluates reasoning and comprehension benchmarks rather than qualitative coding specifically; the directional asymmetry is a strong prior for analysis design, not a measured effect on thematic analysis.
The open question the field has not answered: how much does switch-point transcription error actually change the themes a study produces? Nobody has run that experiment end to end, and until someone does, the conservative practice is verification.
FAQ
What is code-switching in qualitative research?
Code-switching is when a speaker alternates between two or more languages within a single conversation, often within a single sentence. In interviews it typically appears as a dominant "matrix" language carrying embedded words or phrases from another language — commonly English technical or product vocabulary inside a non-English sentence.
Does AI transcription handle code-switching?
Partially, and unevenly. The 2026 ServiceNow-AI benchmark found that top-performing speech-recognition systems take only a small accuracy penalty on code-switched audio while weaker systems degrade substantially. Because the gap between systems widens on mixed-language speech, you should verify your chosen system on your own bilingual pilot recordings rather than relying on a monolingual accuracy comparison.
Should I translate transcripts to English before analysis?
Usually not by default. Evidence from "Lost in the Mix" (Mohamed et al., 2025) indicates that inserting foreign tokens into English text degrades model comprehension, whereas embedding English inside another language often improves it. Analyzing in the participant's matrix language keeps you on the favorable side of that asymmetry. Translate for reporting, not for coding.
What is a switch point and why measure it separately?
A switch point is the boundary where the speaker changes language. Errors cluster in the two-to-three-word window around it, so a corpus-level word error rate can look acceptable while the most meaning-dense passages are wrong. Reporting switch-point accuracy separately surfaces a failure the aggregate number hides.
Is a loanword the same as code-switching?
No. A loanword is a borrowed term that has become part of the matrix language and is used by monolingual speakers of it. A code-switch is a genuine alternation between language systems. The HiKE framework labels the two separately for exactly this reason — conflating them inflates error estimates and adds review work with no quality gain.
How many transcripts should I audit for switch-point accuracy?
There is no published standard, and we will not invent one. A defensible internal practice is to audit a random sample large enough to cover every recruited language pair and every interviewer, and to expand the sample if the first pass surfaces errors that would change a quote or a code. Document what you sampled and why.
Bottom line
Code-switching in multilingual interviews is not a transcription nuisance to be cleaned up later — it is a structural weak point that sits exactly where your richest data lives. The 2026 evidence says the tooling gap is real but uneven, that the direction of language mixing changes how well models cope, and that participants switch languages for reasons worth preserving. None of that is solved by choosing a better model. It is solved by declaring your language pairs, auditing switch points, analyzing in the matrix language, and having a bilingual human verify every quote you publish.
Qualitati supports AI-moderated interviews, AI-moderated focus groups, conversational surveys, and AI thematic analysis in ten languages. Start free with 30 credits — no credit card required — or view transparent pricing to see published per-credit rates.
Last updated: August 21, 2026. This article is an independent editorial summary of publicly available research and product information; competitor and tool claims reflect public sources as of the publication date. Methodology guidance here is a review procedure, not a validated instrument — teams making regulated or publication-bound claims should seek methodological review.