The short version: for English, Spanish and the big European and Asian languages, AI now does the work and a light check is enough. For most of Africa's large languages, AI gives you a good draft that a native speaker needs to check. For a long tail, it still can't hear the language at all. And two things cut across every rating: which provider you use, and who is speaking. The same set of Hindi phone calls can come back with 6% of words wrong or 34%, depending on the engine. The same system can get English near perfect when it's read into a studio microphone, and a third wrong on calls to an advice charity in North East Scotland.

What changed since August

We published the first version of this guide in July and updated it in August. When we checked it again in October, 23 of the 74 ratings had moved, and we'd missed a lot of languages people actually field in. So we rebuilt it against what the providers support today and the newest real-world benchmarks.

• Up. Cantonese, Punjabi, Odia, Assamese and Nepali are now Ready, and so is Levantine Arabic, though that rests on one benchmark. Bhojpuri, Luganda, Wolof and Nigerian Pidgin moved from Hard to Workable. Pidgin finally has a commercial speech model (Intron Sahara v2, March 2026). Translation into isiZulu, Kinyarwanda and Javanese is now Strong, and Xitsonga, siSwati and Kirundi moved up to Moderate.
• Down. Hausa and Egyptian Arabic dropped to Workable once we looked at real recordings rather than clean test audio. Moroccan Darija, Sudanese Arabic and Kurmanji Kurdish dropped to Hard. Hebrew is now a borderline Workable. Translation into Wolof and Fula dropped to Basic.
• New. 56 languages added, including Turkish, Malay, Gulf and Iraqi Arabic, Pashto, Dari, Central Asian languages, 14 more sub-Saharan African languages, Tamazight and Kabyle, Haitian Creole, and indigenous languages of Latin America such as Quechua, Aymara, Guarani, Nahuatl and the main Mayan languages.
• Advice we've changed. We used to suggest transcribing first and letting an LLM clean up mixed-language text. A 2026 study found that made Hindi transcripts worse. We also used to point to Meta's MMS model for the long tail. Meta has since replaced it with Omnilingual ASR (November 2025, open source, 1,600+ languages), though it still leaves out Sesotho, siSwati, Tshivenda and isiNdebele.
• New sections. Which provider wins where, which model translates best, how accents change the picture inside “ready” languages, and the language mix inside 23 markets.

How the ratings work

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 another language. Each language is rated on the best tool available on the market, not on any one provider, and we apply the same rules to every language.

• ●
  Ready (transcription). At least two commercial providers, and either ElevenLabs rates its own accuracy at 10% word error or better, or the best real-world benchmark is around 15% or better.
• ◐
  Workable. At least one current commercial provider, including regional specialists such as Intron, Sarvam and Lelapa, with real-world error of roughly 15 to 35%, or a listing with no published accuracy.
• ○
  Hard. No current commercial speech-to-text, or best evidence worse than about 35 to 40% of words wrong.
• ●
  Strong / Moderate / Basic (translation). Based on how closely frontier models (GPT, Claude and Gemini) match a reference translation into the language, on the AI Language Proficiency Monitor's 0 to 100 chrF score: Strong at 57 or above, Moderate from about 42 to 56 or listed by Google Translate or DeepL with no benchmark, Basic below that. We rate up where the score is known to undercount, such as Ge'ez script (Amharic, Tigrinya) and Yoruba tone marks.

The language table

All 130 languages, grouped by region. The last column tells you what to do in fieldwork and the evidence behind the rating. Moved marks a rating that changed since August, New a language we added. Speaker counts are broad public estimates.

● Ready AI does the job, light spot-check.  ◐ Workable AI drafts, a native speaker checks.  ○ Hard Human-led, AI assists.

About 36 of these ratings rest on thin evidence, mostly newer African languages where all we have is a provider listing or an open-model test. They're marked in the table. The ratings move quickly, so check the date on anything you rely on.

Languages inside each country

A language rating only goes so far, because nobody runs a study in “isiZulu”. They run it in South Africa, where English is spoken at home by fewer than one in ten people. So we mapped the home languages of 23 markets, using censuses and statistics-office surveys, and coloured each language by its transcription rating.

Share of people by home language, coloured by AI transcription rating. Sources: national censuses and statistics-office surveys 2007 to 2025; Afrobarometer Round 10 (2024) for Nigeria, Kenya, Ghana, Uganda and Tanzania.

The split is stark. In Brazil, the UK, the US, Mexico and Spain, more than nine in ten people speak a home language AI transcribes well. In Ethiopia, Uganda, Nigeria, Ghana and Pakistan, fewer than one in ten do. South Africa sits at about one in five, because English and Afrikaans are Ready but isiZulu, isiXhosa, Sepedi, Setswana and Sesotho are only Workable.

On the country map below, each country takes the weakest rating among its main spoken languages: the most widely spoken one, any home language of roughly a fifth of people, and any shared language routinely used in fieldwork. Of the 22 countries where a main language is still Hard, 20 are in Africa.

Each country coloured by the weakest transcription rating among its main spoken languages. Grey: the most widely spoken language isn't rated.

Which provider for which language

“Use AI” isn't one decision. The engines differ a lot, and outside English the difference is not a few percent, it can be several times over. Here is what the independent and semi-independent benchmarks from 2025 and 2026 show when they test the big providers on the same real audio.

Share of words (or characters, for Korean, Japanese and Cantonese) transcribed wrong. Sources: Artificial Analysis (English, Oct 2026), Voice of India (Hindi, Telugu), GigaSpeechBench (others).

What the head-to-heads show

• On English, it barely matters. On the Artificial Analysis leaderboard, eight big providers' newest models all get between about 2% and 5% of words wrong.
• On Indian phone calls, it matters a lot. In Voice of India, an independent benchmark of real phone conversations in 15 languages, Sarvam (an Indian specialist) was the best commercial system on 13 of them. Among the big global providers, Google's Gemini 3 Pro led on 11, but Amazon Transcribe won Telugu, Kannada and Odia. OpenAI's GPT-4o Transcribe got 34% of Hindi words wrong against Gemini's 6%, and AssemblyAI's output was unusable in six languages.
• On Arabic, it depends on the country. In GigaSpeechBench (real YouTube audio, co-written by Alibaba), Google's Chirp 3 was the best Western provider on Saudi Arabic, Google's Gemini on Syrian and Egyptian, and Microsoft Azure on Iraqi. Two Google products swap places from one country to the next.
• On East and Southeast Asian languages, there's no pattern. Among Western providers, ElevenLabs led on Korean and Thai, Google's Chirp 3 on Vietnamese and Indonesian, and Microsoft on Japanese. On Cantonese, Microsoft got 12% of characters wrong and Google's Chirp 3 got 48%. (Alibaba co-wrote this benchmark, and its own models were best on several of these languages.)
• On mixed African-English speech, everyone struggles. In AfriSwitch, Gemini and ElevenLabs split the 12 languages six each, averaging 55 to 56% of words wrong. Intron's own model averaged 36%, but Intron also built the benchmark.
• On African-accented English, Microsoft, Amazon and Google were within about two points of each other on average (22 to 24%), with the winner changing by country. Again an Intron benchmark, where Intron's model was well ahead.

Two caveats. Model versions change every few months, so any ranking is a snapshot. And the newest models (Gemini 3.5 Transcribe, GPT Transcribe, MAI-Transcribe-2) only have vendor figures outside English so far. The practical answer is to shortlist two or three providers for each language and test them on 20 or 30 of your own voice notes before you commit.

Which model translates best

Translation is closer, but the models still aren't interchangeable. By our calculation from the AI Language Proficiency Monitor's published results, which score frontier models on translation into 194 languages, Gemini 3.1 Pro averaged 52.4, GPT-5.5 50.9 and Claude Opus 4.8 49.6 (the earlier Claude Opus 4.7 scored 51.1). Gemini came top in 124 languages, GPT in 36 and Claude in 35. Each language is scored on only 10 sentences, so gaps of a few points on a single language don't mean much.

Two things matter more for research than the leaderboard. Benchmarks use formal text, and in WMT25 transcribed speech was the hardest material to translate, likely because of errors carried over from the transcription. And automatic scores and human judges don't always agree, as the Hausa study shows. For open-ended answers, have a native speaker read a sample of translations alongside the originals, especially where people use slang or switch languages.

Accents inside “ready” languages

“Ready” is a rating for a whole language, and it hides a lot. On standard test sets the best systems get about 2% of English words wrong. On real people, the picture changes depending on where they're from and who they are.

Each line compares one system on two groups of speakers in the same study. Sources: VarDial 2025; Markl, FAccT 2022; Koenecke et al., PNAS 2020; Harris et al., EMNLP Findings 2024.

• Scotland. On real Scottish speech in GigaSpeechBench, the best system tested got 24% of words wrong (an Alibaba model, in a benchmark Alibaba co-wrote). Microsoft Azure got 28%, and ElevenLabs, Gemini and GPT-4o were between 34% and 44%. Whisper went from 4% on its baseline to 22 to 34% on Scottish housing and charity calls.
• Northern England and Northern Ireland. On Newcastle speech, the best of four systems averaged 32%, with Google and Deepgram at 40 to 60%. In an older study of nine British cities, Amazon Transcribe was 9 points worse in Belfast and Bradford than in Cambridge.
• Black American speakers. In the reference study from 2020, five commercial systems averaged 35% against 19% for white speakers, and 23% of clips were unusable against 1.6%. The systems are older, and we found no repeat of the test on today's models.
• Spanglish and Chicano English. An open model got 42% and 38% of words wrong, against 20.5% on standard American English. We found no independent test of today's commercial providers on real Spanglish conversation.
• Varieties of the same language. A model tuned on European Portuguese made 27% errors on Brazilian Portuguese against 12.5% on its own variety. Southern Spain accents ran up to 8 points worse than northern ones. Québec French came out at 8 to 10%, but standard French benchmarks didn't predict which model would do best.

The home languages of UK and US samples matter too. Only 78.4% of Londoners have English as their main language (91.1% across England and Wales). The next biggest languages nationally are Polish, Romanian, Panjabi and Urdu. Polish and Romanian are Ready. Panjabi and Urdu are Ready only on evidence from Indian audio (Sarvam and Gemini). ElevenLabs rates Urdu at 25 to 50% of words wrong, Google's Chirp 3 has no Pakistani Urdu locale, and we found no test on UK speakers. In the US, Spanish is spoken at home by 13.9% of people, followed by Chinese, Tagalog, Vietnamese and Arabic, all Ready, while Haitian Creole, with about a million speakers, is only Workable.

Markets we haven't mapped

Our 23-market map leaves out some markets that come up in almost every multi-country brief. Here is where their main languages stand.

Code-switching

Real people don't speak one clean language at a time. A respondent in Nairobi slides between Swahili, English and Sheng in a single thought. In Johannesburg it's English threaded through isiZulu. In Lagos, English, Yoruba and Pidgin. In Manila, Taglish. In India, Hinglish.

Most speech models assume one language per sentence, and the gap shows. AfriSwitch, a 2026 benchmark of 61 hours of real conversation mixing English with 16 African languages and varieties, found no system below 24% on any of the 12 languages benchmarked so far. The best, Intron's own Sahara V2.5, averaged 36% across the 12 languages in its main results; Meta's Omnilingual, Google's Gemini and ElevenLabs averaged 51 to 56%. In the Philippines, where Taglish is everyday speech, the best real-world results for Tagalog were still around a quarter of words wrong.

Two things help. Use a speech model trained on mixed speech where one exists (Sarvam, for example, supports Hinglish). And don't assume an LLM can tidy it up afterwards: asking GPT-4o-mini to correct Hindi transcripts without examples took word error from 18% to 25%. Keep the original audio, and have a native speaker check the parts where people switch.

Recording conditions

Benchmark accuracy is measured on clean recordings. Your data is a voice note recorded on a cheap phone in a taxi or a market. Expect error rates roughly 1.5 to 2 times worse than the clean figures for single-language speech. On real Indian phone calls, the noisiest quarter of recordings scored about 1.65 to 1.75 times worse than the cleanest; on African conversational audio the gap ranged from 1.3 to 2.8 times.

The cause matters. WhatsApp's own audio compression (Opus) barely moves accuracy. Old-style 8 kHz phone lines do. And competing voices in the background do the most damage of all. So the cheapest improvement isn't a better model, it's asking people to record somewhere quiet, away from the TV and other conversations, and testing models on your own pilot audio rather than their published numbers.

So, AI or a team?

Put it together and it isn't “AI” or “humans”. It's a split that depends on the language, the provider and the people.

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

That's how research happens in people's own words at scale without quietly shipping bad data. “We can only do this in English” is no longer true for most large languages. Which languages, in which mode, with how much human review, is now a decision worth making on purpose.

Frequently asked questions