Dataset Card for Vuk'uzenzele isiXhosa Speech Dataset (ViXSD)

Dataset Description

Dataset Summary

Vuk'uzenzele isiXhosa Speech Dataset (ViXSD) contains scripted narration of the Vuk’uzenzele South African Multilingual Corpus. ViXSD contains read speech from native speakers accompanied with rich metadata on speaker demographic and linguistic distribution. ViXSD consists of 395 stereo audio recordings and corresponding transcriptions derived from the Vuk’uzenzele South African Multilingual Corpus. It contains a total of 10 hours of narrated speech isiXhosa narrated by 8 speakers (4 male, 4 female) with approximately 39,000 words. We split the data into train, dev and test split for ease of use.

Paper: coming soon
Point of Contact: Lelapa AI
Data Sheet: Datasheet(pdf)
License: Esethu License

Languages

isiXhosa

How to use

The datasets library allows you to load and pre-process your dataset in pure Python. The dataset can be downloaded and prepared in one call to your local drive by using the load_dataset function.

For example:

from datasets import load_dataset, DatasetDict
visxd_train = load_dataset("lelapa/isixhosa_community_dataset", split = 'train')

Using the datasets library, you can also stream the dataset on-the-fly by adding a streaming=True argument to the load_dataset function call. Loading a dataset in streaming mode loads individual samples of the dataset at a time, rather than downloading the entire dataset to disk.

from datasets import load_dataset
visxd_train = load_dataset("lelapa/isixhosa_community_dataset", split = 'train', streaming=True)

Example scripts

Fine-tune your own Automatic Speech Recognition models on ViXSD with mms - here.

Dataset Structure

Data Instances

A typical data point comprises the audio_paths to the audio file and its transcript. Additional fields include 'idx', 'user_ids', 'accent', 'age_group', 'country', 'nchars', 'audio_ids', 'duration', 'origin', 'domain', 'split', 'gender', 'title', 'author', 'location_born', 'location_childhood', 'location_current', 'first spoken language', 'multi-lingualism_and_confidence'.

{'idx': 2,
 'user_ids': 'xho_reader_003',
 'accent': 'xho',
 'age_group': '30 - 39',
 'country': 'ZA',
 'transcript': 'URakwena uthi le mali ichithwa kule nkqubo iyongezwa rhoqo ngonyaka kulungiselelwa umba wokunyuka kwamaxabiso ezinto, ke ngoko imali ebekelwe i-NSNP yowama-2017/18 imalunga neebhiliyoni eziyi-6.4 zeerandi. I-NSNP ilicebo lexesha elifutshane lokugxotha ikati eziko kwaye inegalelo elihle kakhulu kubomi babafundi.Abantwana abaninzi baphuma kumakhaya angathathi ntweni, angenamali yokuthenga into esiwa phantsi kwempumlo, kwaye abanye abantwana oku kutya bakufumana esikolweni, yiyo yodwa abayityayo ngosuku. Ukutya abakufumana esikolweni kuyabanceda abafundi kwizifundo zabo, bakwazi ukuhlala bemamele bethatha nenxaxheba kwizifundo zabo, utshilo. URakwena wongeze ngelithi amanani abantwana abaya esikolweni enyukile futhi abantwana basihamba kakuhle isikolo ngenxa yale nkqubo yokutyiswa kwabantwana ezikolweni. ',
 'nchars': 813,
 'audio_ids': 'cds_xho_001_2',
 'audio_paths': {'path': 'cds_xho_001_2.wav',
  'array': array([ 0.00334167,  0.00346375,  0.003479  , ..., -0.00041199,
         -0.00045776, -0.00048828]),
  'sampling_rate': 44100},
 'duration': '0:00:55',
 'origin': "Vuk'uzenzele-2018-01-ed1",
 'domain': 'General',
 'split': 'train',
 'gender': 'Female',
 'title': 'Abantwana abahluthiyo baqhuba kakuhle kwizifundo zabo',
 'author': 'More Matshediso',
 'location_born': 'Eastern Cape, South Africa',
 'location_childhood': 'Western Cape, South Africa',
 'location_current': 'Western Cape, South Africa',
 'first spoken language': 'isiXhosa',
 'multi-lingualism_and_confidence': 'English - 10; isiXhosa - 10; isiZulu - 9'}

Data Fields

idx (int64): An id for the audio file

user_ids (string): An id for the speaker

accent (string): Accent of the speaker 'xho'

age_group (string): Age group of the speaker

country (string): The locale of the speaker

transcript (string): The sentence the user was prompted to speak

audio_ids (string): Unique name of the audio file

audio_paths (audio): The path to the audio file

duration (string): Duration of the audio file (00:00:00)

origin (string): Reference to Vuk'uzenzele article

domain (string): Domain of the transcript ('General')

split (string): Train/Dev/Test

gender (string): The gender of the speaker

title (string): Title of the Vuk'uzenzele article

author (string): Author of the Vuk'uzenzele article

location_childhood (string): Author of the Vuk'uzenzele article

location_current (string): Author of the Vuk'uzenzele article

first spoken language (string): Author of the Vuk'uzenzele article

multi-lingualism_and_confidence (string): Author of the Vuk'uzenzele article

Data Splits

Split	Participants	Gender	Age	Total Duration	Size	Avg Duration	SNR
		male/female	<29/<39/<49	h:mm:ss	#transcriptions	mm:ss per file	dB
train	4	2/2	1/3/0	7:47:01	304	01:32	-0.01
dev	2	1/1	0/2/0	1:33:00	65	01:26	0.44
test	2	1/1	0/1/1	0:40:13	26	01:33	-0.15

Dataset Creation

The dataset contains scripted narration of the Vuk’uzenzele South African Multilingual Corpus.

The audios were narrated by equal representation of male and females with education levels ranging from NQF Level 5 to Bachelors Degrees. The speakers are within the age range of 18-40 years, with 12.5% being between 18–29 years, 75.5% between 30–39 years and 12.5% being 40 years. We focus primarily on isiXhosa with occasional code-switching where the script included non-isiXhosa terms (e.g., organizational names in English). We designed our selection process to reflect inclusivity within isiXhosa-speaking South African society. Where a person was born, has grown up and lives has a big impact on their accent as well as the vocabulary they will use when speaking their language.

Personal and Sensitive Information

The dataset consists of people who have donated their voice online. You agree to not attempt to determine the identity of speakers in the Common Voice dataset.

Considerations for Using the Data

Social Impact of Dataset

isiXhosa, a Bantu language spoken by over 9 million people in South Africa. isiXhosa is notably underserved despite its sizable population.

Discussion of Biases

The creation of this isiXhosa dataset necessitates careful consideration of ethical issues, particularly regarding data privacy, cultural representation, and linguistic bias. The dataset is derived from publicly available material, ensuring compliance with copyright and data protection regulations. Personally identifiable information (PII) is either absent or anonymised to safeguard individual privacy.

Given the dataset’s focus on news type articles, it may exhibit thematic and linguistic biases. Efforts have been made to preserve dialectal diversity and cultural authenticity, yet idiomatic expressions and culturally specific references may require careful interpretation in downstream NLP applications. Researchers should be cognisant of potential stereotypical or skewed linguistic representations when applying models trained on this corpus.

Ethical considerations

Additional Information

Licensing Information

Esethu License.

The dataset was developed with an ethical approach to data licensing as a central consideration in the release process. The creators of the dataset are committed to ensuring that the primary beneficiaries of the dataset are: (a) the contributors, in this case, the isiXhosa community, and (b) the broader African language technology ecosystem.

We have developed a novel license for this dataset. This license consists of two components: a commercial/proprietary license and an open license. In essence, the license does not restrict research use and permits commercial use by African entities. However, non-African commercial entities are required to pay a licensing fee.

Proceeds from the dataset are legally mandated to be reinvested in the creation of more isiXhosa data through a local data creation partner, which will be released under the same license. This ensures a circular system that not only promotes the generation of additional isiXhosa data, benefiting the isiXhosa community through language technology opportunities, but also guarantees that the revenue generated from the dataset supports fairly paid work for isiXhosa speakers. By ensuring that the data is accessible to African commercial entities, we aim to foster innovation within the African language ecosystem while preserving opportunities for research on the language.

Citation Information

@misc{rajab2025esethuframeworkreimaginingsustainable,
      title={The Esethu Framework: Reimagining Sustainable Dataset Governance and Curation for Low-Resource Languages}, 
      author={Jenalea Rajab and Anuoluwapo Aremu and Everlyn Asiko Chimoto and Dale Dunbar and Graham Morrissey and Fadel Thior and Luandrie Potgieter and Jessico Ojo and Atnafu Lambebo Tonja and Maushami Chetty and Onyothi Nekoto and Pelonomi Moiloa and Jade Abbott and Vukosi Marivate and Benjamin Rosman},
      year={2025},
      eprint={2502.15916},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2502.15916}, 
}

Datasets:

lelapa
/

Vukuzenzele_isiXhosa_Speech_Dataset_ViXSD

You need to agree to share your contact information to access this dataset