South African speech data: isiZulu, isiXhosa and Sesotho
License 2,000 hours of South African English conversation off the shelf, or have isiZulu, isiXhosa and Sesotho speech data recorded to your spec by native speakers. Other South African languages are available on request, with feasibility confirmed per project.
South African English, off the shelf
2,000 hours of spontaneous two-person conversation from 200 native speakers recorded in South Africa, with one channel per speaker and time-aligned transcripts, at $60 per hour of audio. See the dataset and request samples.
isiZulu, isiXhosa and Sesotho, recorded to spec
Native speakers recruited to your quotas for age, gender and region, recording scripted prompts or spontaneous conversation on the devices and channels you specify. Transcripts follow your convention, and consent is documented for every speaker.
What makes these languages hard for speech recognition
Public speech data for these languages is limited, and most of it is read or prompted speech rather than conversation.
Click consonants
isiXhosa and isiZulu use click consonants, written c, q and x. Models trained mostly on other languages have heard few of them.
Tone
All three are tone languages, and standard spelling does not mark tone, so two words spelled the same can differ in pitch and meaning.
How words are written
isiZulu and isiXhosa are written conjunctively: prefixes and concords join the verb, so ngiyakuthanda (“I love you”) is one isiZulu word. Vocabularies grow large, and one wrong prefix counts as a whole word error. Sesotho writes the same kinds of pieces as separate words.
Code-switching
Speakers often switch between their home language and English, sometimes inside a sentence. Transcripts need rules for mixed-language speech, and training data needs to contain it.
Built for products that serve South African users
Contact centers and voice agents
Callers may switch between English and their home language in one call. We can record role-played customer service calls in your scenarios, with one channel per speaker.
Speech recognition
Training and evaluation data with the accents, code-switching and recording conditions your users produce, with held-out test sets if you need them.
Text-to-speech
Scripted recordings from selected native speakers under consistent recording conditions, with prompts that cover clicks and the vocabulary your product says aloud.
Frequently asked questions
Do you have isiZulu, isiXhosa or Sesotho speech datasets?
Not off the shelf. isiZulu, isiXhosa and Sesotho speech data is recorded to your spec by native speakers, scripted or spontaneous, with time-aligned transcripts and documented consent. Our off-the-shelf South African dataset is South African English.
Can recordings include code-switching with English?
Yes. Spontaneous conversations can keep switches between English and the home language as they happen, and transcripts can tag code-switches as your guideline specifies.
Which other South African languages can you collect?
Other South African languages, such as Afrikaans, Sepedi or Setswana, can be requested. We confirm recruitment feasibility for each language before quoting.
Tell us which South African languages you need
Share languages, hours, speaker profiles, recording conditions and timeline. We’ll come back within one business day with a feasibility assessment and a quote.