For most of the languages you care about, AI now does the heavy lifting and a small human layer catches what it misses. For a stubborn minority, that ratio flips and people still need to lead. The whole game is knowing which bucket your language falls into before you start, which is exactly what the table below is for.

The language table

Two separate ratings, because they're two separate jobs. Transcription turns a voice note into text in the same language. Translation turns that text into English. Transcription is the harder job and where nearly all the difficulty lives; translation runs roughly a tier ahead almost everywhere.

● Ready AI does the job, light spot-check only.  ◐ Workable AI drafts, a human closes the gap.  ○ Hard Human-led, AI assists.

Grouped by region, with the languages most relevant to emerging-market fieldwork first. "Approach" is the setup we'd actually reach for.

One thing about these ratings: they move. Investment in African and emerging-market languages has accelerated sharply, so a language sitting in "Hard" today can be "Workable" within a year. The floor is rising fast, and faster for the languages with the most speakers.

Which model for which job

"Use AI" isn't one decision. The engines differ enormously, and routing to the wrong one is the most common reason a multilingual study underperforms. A serious operation sends each language to its best model rather than forcing everything through a single provider, which is the logic behind the "recommended approach" column above.

No single model wins everywhere, and the same is true for translation. Google Translate has the widest coverage, Meta's open NLLB is strong on low-resource languages, and modern general-purpose LLMs now match or beat dedicated translation engines on the top languages and handle slang and mixed text more gracefully, while still weakening on the minority tail.

Code-switching is the thing that breaks everything

Here's the failure mode that catches people out, because it shows up in no benchmark. Real people don't speak one clean language at a time. A respondent in Nairobi slides between Swahili, English, and Sheng within a single thought. In Johannesburg it's English threaded through isiZulu. In Lagos, English, Yoruba, and Pidgin at once. In India, Hinglish and Tanglish.

Almost every speech model is trained on the assumption of one language per utterance, so code-switching is exactly where they quietly fall apart, and in emerging markets this isn't an edge case. For a large share of urban respondents, mixing languages is simply how they talk.

The silver lining: large language models handle code-switched text far better than speech engines handle code-switched audio, since they're trained on the messy way people actually write. The winning pattern is a decent transcript first, then an LLM to clean, interpret, and translate the mix, rather than expecting the speech engine to nail it in a single pass. Even so, this remains the single biggest reason to keep a human in the loop on the harder tiers. Code-switching is where the machine is most confidently wrong.

The recording environment matters more than the model

Benchmark accuracy gets measured on clean studio audio. Your data is a voice note recorded on a cheap phone, in a taxi, in a market, with wind and a TV in the background, then squeezed through WhatsApp compression.

Based on our own fieldwork, expect error rates roughly 1.5 to 2 times worse than any benchmark number, purely from the audio conditions. That's not the model failing, it's physics. Two things follow. Denoising before transcription genuinely moves the numbers, which is why models with built-in cleanup have a quiet advantage on field audio. And a little guidance to respondents, find a quiet spot, hold the phone close, improves data quality more cheaply than any model upgrade. The environment is the variable people ignore and then blame the AI for.

So, AI or a team?

Put it together and the answer isn't "AI" or "humans." It's a ratio that shifts with the job.

The old model, a large human team transcribing and translating everything by hand, is now the wrong answer for almost every study. It's slow, it doesn't scale, and it's expensive where it doesn't need to be. But pure AI with no human anywhere is equally wrong for the hard cases, because it fails silently, and you find out when a client questions a finding you can't defend.

The right answer is AI-first, with humans placed exactly where the technology is weak.

Route each language to its best model, clean the audio, let AI carry the volume, and spend human hours on the specific failure points: the code-switching, the low-resource languages, the nuance that carries the insight. That's how research happens in people's own words, at scale, without burning budget or quietly shipping bad data. "We can only do this in English" is no longer a real constraint. Which languages, in which mode, with how much human review, is now a strategy decision, and one worth making on purpose.

Frequently asked questions