Home › Languages we recruit
The languages and accents we are recruiting voices for
Speech data has an availability problem, not a talent problem. The voices exist - they are just not on the platforms buyers look at. We recruit directly, in the languages and accents that datasets keep missing.
Where we recruit
By language family
This is a recruitment list, not a stock catalogue. Speakers are recruited and recorded per project, with consent and licence agreed before anything is captured. If your language is not listed, apply anyway - the list follows demand, and demand moves.
English & English accents
The most recorded language in the world and still the most requested, because general models trained on neutral American and southern British English fail on everyone else. We recruit specifically for the accents that are missing: US regional accents from every state, Scottish, Welsh, Irish, northern English, West Midlands, Multicultural London English, Indian English, Nigerian and Ghanaian English, Caribbean English, Singaporean and Malaysian English, South African, Canadian, Australian and New Zealand English.
Now casting: US voices for an AI telephony project.
Western European
Dutch and Flemish, French, Belgian and Canadian French, German, Austrian and Swiss German, Spanish and Latin American Spanish, Portuguese and Brazilian Portuguese, Italian, Catalan, Basque, Galician, Luxembourgish, Maltese, Welsh, Scottish Gaelic, Irish, Danish, Swedish, Norwegian, Finnish, Icelandic, Faroese, Greek. Regional variants matter here as much as the language itself - Flemish is not Netherlands Dutch, and models trained as if it were fail in Belgium.
Central & Eastern European
Polish, Czech, Slovak, Hungarian, Romanian, Bulgarian, Croatian, Serbian, Bosnian, Slovenian, Macedonian, Albanian, Ukrainian, Belarusian, Russian, Lithuanian, Latvian, Estonian, plus Georgian, Armenian, Azerbaijani, Kazakh and Uzbek. Consistently under-supplied relative to demand, particularly for spontaneous and conversational speech rather than read prompts.
South Asian
Hindi, Urdu, Bengali, Punjabi, Gujarati, Marathi, Tamil, Telugu, Kannada, Malayalam, Odia, Assamese, Sinhala, Nepali. Also code-switched speech - Hinglish and its equivalents - which is how hundreds of millions of people actually talk and which almost no dataset captures properly.
Southeast & East Asian
Tagalog and Filipino, Cebuano, Bahasa Indonesia, Javanese, Bahasa Malaysia, Thai, Lao, Vietnamese, Khmer, Burmese, Mandarin (including Taiwanese Mandarin), Cantonese, Japanese, Korean, Mongolian. Tonal languages need speakers recorded to a standard that preserves pitch contour cleanly, which rules out a lot of crowdsourced audio.
Middle East & North Africa
Modern Standard Arabic plus the dialects that people actually speak - Egyptian, Levantine, Gulf, Iraqi, Maghrebi - along with Hebrew, Turkish, Farsi, Dari, Pashto, Kurdish and Amazigh (Berber). Dialectal Arabic is one of the widest gaps between what models are trained on and what users say to them.
Sub-Saharan African
Swahili, Amharic, Yoruba, Igbo, Hausa, Zulu, Xhosa, Afrikaans, Somali, Oromo, Tigrinya, Shona, Kinyarwanda, Lingala, Wolof, Twi and Akan. Among the most under-represented language families in commercial speech data, and among the fastest-growing in demand.
Why accent coverage is a technical requirement
A speech model does not fail gracefully on an accent it has never heard. It fails loudly, and it fails on the users least likely to be asked for feedback. Word error rates on under-represented accents routinely run several times higher than on the accents that dominate training data, and the gap does not close by adding more of the same voices.
Under the EU AI Act, data governance for high-risk systems includes examining datasets for gaps and bias. Accent and dialect coverage stops being a nice-to-have at that point and becomes something you have to be able to evidence. We build recruitment around that from the start - see how Article 10 shapes the work.
For voice talent
If you speak a language or accent on this page, you are more in demand than the general voiceover market suggests, and the scarcer your accent the truer that is. You do not need a showreel or an agent. Starting with no experience explains the route, and what the work pays covers the money and the licence.
For AI companies
Tell us the languages, the accent spread, the speaker demographics and the hours. We recruit against the brief rather than filtering whoever happens to already be on a platform, which is why the coverage looks different. Multilingual speech data covers how those projects run.
Your language is probably on someone’s shortage list
Apply once. We come to you when a project matches your language and accent.