Loading datasets...
Release Date: 11/4/2026
Format: WAV
Size: 1.40 GB
The dataset was collected in Kisumu, Nyanza region of Kenya, during a pilot conducted from May to June 2026 using the TWB Voice platform. The project was designed to address the limited availability of high-quality spontaneous speech resources for Dholuo, a low-resource Kenyan language, particularly for the development of Automatic Speech Recognition (ASR) systems supporting agricultural advisory services. The dataset contains approximately 20 hours of spontaneous Dholuo speech and 20 hours of corresponding transcriptions, organised as audio–text–metadata triples. Each dataset item consists of a natural spoken response, its validated transcription, and associated metadata describing characteristics such as the speaker's age, gender, language variant, and recording conditions. The speech was intentionally collected as spontaneous, unscripted responses rather than read-aloud material, making the dataset suitable for research on real-world spoken Dholuo. Data collection involved 26 Dholuo-speaking contributors, including 21 participants recruited and trained through an in-person workshop and five additional contributors recruited and onboarded online during the pilot. Contributors were selected to provide variation in age, gender, and dialect, with the aim of representing everyday Dholuo rather than a single standardised variety. The intended application domain was agricultural advisory, and contributors responded to prompts covering crop farming, livestock management, poultry keeping, fish farming, soil health, climate adaptation, pest control, and market access. The project developed approximately 2,000 text-based prompts and 1,000 image-based prompts to elicit relevant spontaneous speech. The collection followed a structured quality-control pipeline. Contributors recorded responses in quiet environments, with individual statements generally lasting 15–60 seconds, and were encouraged to listen to their recordings before submission. Submitted recordings underwent voice validation, where samples containing multiple speakers, excessive background noise, off-topic content, no human voice, or speech in a language other than Dholuo could be rejected. Accepted recordings were transcribed through a two-level transcription process: an initial transcription followed by independent checking for accuracy, omissions, and consistency with the project style guide. The transcription guidelines were designed to preserve characteristics of naturally occurring Dholuo speech while maintaining consistency. Repeated words and meaningful filler words were retained, recoverable truncated words were completed using contextual information, and instances of English or Kiswahili code-switching were explicitly marked using [cs]... [cs] tags. These annotations provide additional information that may be useful for studies of multilingual and code-switched speech.
Licensing
Nwulite Obodo Open Data Licence 1.0 (NOODL-1.0)
https://licensingafricandatasets.com/nwulite-obodo-licensePrivacy and protection of contributors: The dataset contains recordings of human voices, which may carry identifiable characteristics even when names, email addresses, and other direct identifiers have been removed. Users must not attempt to identify individual contributors, link recordings to external personal records, infer identities, or disclose information that could compromise participants' privacy. The dataset must not be used for voice identification, speaker re-identification, biometric profiling, surveillance, or other activities that could expose contributors to harm without explicit authorisation from the relevant data providers and any other required approvals. Users must take appropriate measures to protect downloaded files and derived datasets against unauthorised access, disclosure, and misuse.
Respect for informed consent and community interests: The dataset was collected from adult Dholuo-speaking contributors who participated under an informed-consent process. Its use should remain consistent with the purposes, permissions, and limitations communicated during collection and established by the applicable licence and data-sharing arrangements. Researchers and developers should respect the rights of contributors, including applicable withdrawal arrangements, and should direct requests concerning consent, withdrawal, corrections, or other participant-related matters to the authorised dataset custodians. The resource must not be used in ways that discriminate against, stigmatise, exploit, or misrepresent Dholuo-speaking communities.
Permitted research and technological applications: The primary intended use is the development and evaluation of automatic speech recognition (ASR), speech-to-text systems, and voice-enabled agricultural advisory services for Dholuo speakers. Other suitable applications may include low-resource language research, speech processing, multilingual language technology, code-switching analysis, agricultural information access, and the evaluation of language models. Any commercial use, redistribution, incorporation into third-party products, or use for purposes beyond the licence permissions must be assessed against the NOODL-1.0 terms and any applicable restrictions. Users must not assume that research-oriented collection automatically authorises unrestricted commercial exploitation.
Restrictions on harmful or unrelated use: The dataset should not be used to develop deceptive voice systems, generate misleading speech attributed to real contributors, impersonate individuals, or facilitate fraud, harassment, political manipulation, or other harmful activities. It should not be used to make consequential decisions about individuals, such as employment, creditworthiness, eligibility for services, or law-enforcement targeting. Applications involving sensitive profiling or high-impact decisions require particular caution and must not proceed where they conflict with the licence, participant permissions, applicable law, or community protections.
Transcription, code-switching, and data integrity: The recordings and transcriptions represent spontaneous Dholuo speech elicited through agricultural text and image prompts. They may contain dialectal variation, repetitions, meaningful fillers, naturally occurring disfluencies, and English or Kiswahili code-switching marked using [cs]... [cs] tags. Users should preserve these features where relevant to their research objectives and document any normalisation, correction, filtering, or transformation applied to the data. Transcriptions should not be treated as universally standardised representations of all Dholuo varieties. Automated systems trained on this resource should be evaluated for transcription accuracy, dialect coverage, code-switching performance, and potential bias.
Limitations of the pilot dataset: The dataset originated from a pilot involving contributors recruited through in-person and online activities in the Kisumu/Nyanza region of Kenya. It may not represent all Dholuo-speaking populations, dialects, age groups, recording environments, or agricultural practices. The currently published repository contains a limited pilot release, and final approved dataset counts, total hours, and detailed demographic distributions should be confirmed with the custodians before use. Users must not make unsupported claims about representativeness, completeness, or performance across all Dholuo speakers. Agricultural content should also be checked for accuracy and local relevance before being used to provide practical farming advice.
Attribution, redistribution, and derivative datasets: Users must comply with the attribution, notice, sharing, and redistribution requirements stipulated by NOODL-1.0 and any accompanying terms. Where permitted, publications and derivative resources should acknowledge the dataset, the Dhonam Project, Utavu AI Foundation, and CLEAR Global, as appropriate under the applicable conditions. Derived datasets, model checkpoints, transcriptions, and redistributed recordings must be handled according to the governing licence and data-use restrictions. Users should not remove licence notices, misrepresent the source of the resource, or imply that the data providers endorse a resulting product without permission.
Data security and responsible publication: Users must store and transfer the dataset securely, restrict access to authorised personnel, and avoid publishing raw recordings, metadata, or derivative materials that could reveal personal information unless the relevant permissions explicitly allow such publication. Suspected privacy incidents, unauthorised disclosure, or concerns about the dataset should be reported to the authorised custodians. Questions about permissions, licence interpretation, withdrawal requests, or proposed uses outside the stated scope should be resolved with the data providers before proceeding.
These restrictions are intended to promote responsible access, privacy protection, respect for community interests, and sustainable development of Dholuo speech technologies. The applicable NOODL-1.0 licence and the dataset providers' confirmed access conditions remain authoritative where they differ from this summary.
Any attempt to clone the voice or train models that imitate the speakers in this dataset is forbidden It is forbidden to use this dataset to train chatbots or large language models You agree not to attempt to determine the identity of speakers in this dataset
Ethical Review
Ethical data collection was treated as a core component of the project. Contributors were required to be Kenyan citizens aged 18 years or older, and informed consent was explained orally as well as provided in written form. Voice data was pseudonymised before transfer, participants retained the right to withdraw, and data transfers between the European Union and Kenya were governed through appropriate data-sharing arrangements. The dataset is intended for publication under the NOODL licence, with access governed by data-use terms. Published voice data is anonymised, with personally identifying information such as names and email addresses excluded.
Intended Use
This dataset is intended for use in creating automatic speech recognition systems and Speech to text synthesis
Enough information about the audio and their transcripts have been given in the description.
