A search box, a voice interface and an automatic captioning tool all make assumptions about language. They need ways to recognize words, interpret patterns and present useful output. When those foundations are stronger for some languages than others, the difference can shape who finds technology convenient and who must repeatedly adapt to it. South African language technology begins with a practical question: what information and expertise are needed for digital systems to work well in the languages people actually use?
The South African Centre for Digital Language Resources, known as SADiLaR, supports research and development across South Africa’s official languages. Its work includes digital resources, research infrastructure and projects connecting language with technology and the humanities.[1] The center offers a useful way to understand an often overlooked form of advancement: carefully prepared data and tools that other researchers and developers can build on.
Language support is more than a translated menu
A product can display a translated welcome screen while still struggling with the user’s language in search, spelling or speech. Meaningful support reaches into the functions that make the product useful. It includes how words are represented, how variation is handled and whether the system can recognize ordinary use rather than a narrowly selected demonstration.
For a hypothetical municipal information service, translating the homepage would be only one step. Users might search with different word forms, switch languages within a question or use a familiar local term that is absent from the official document. The service needs a way to connect those expressions with the relevant information.
This makes language work an ongoing design responsibility. A team should ask which tasks users need to complete and where communication fails. The answer may involve better text, a revised search resource, human assistance or a different interface. Adding a language name to a list of supported options does not by itself establish that the task can be completed successfully.
A language resource is reusable groundwork
A digital language resource can contain text, speech recordings, annotations, dictionaries or other material prepared for research and development. Its value depends on the purpose it serves and the quality of its documentation. A large collection without a clear description may be harder to use responsibly than a smaller, carefully organized one.
SADiLaR’s repository guidance distinguishes an index of resource descriptions from a catalogue containing downloadable resources, and also describes a student data repository.[2] This is a useful practical distinction. Finding a record does not always mean the associated data can be downloaded immediately or used without checking conditions.
Researchers should examine the resource’s language, format, license and intended scope before deciding whether it fits a project. A collection prepared for one kind of study may not represent the material a commercial interface will encounter. Reuse becomes more productive when the original context travels with the files rather than being lost after download.
Speech data needs people as well as microphones
Speech technology depends on recorded examples, but recording alone does not create a useful dataset. Teams must decide who is represented, what is said, how audio is captured and how the accompanying information is checked. The choices influence what the resulting system can learn and how it should be evaluated.
A hypothetical collection made entirely in a quiet studio may be useful for a defined task but may not reflect the noise of a taxi rank or a shared office. A collection dominated by one group of speakers may also leave important variation underrepresented. The appropriate response is to document the scope and test the intended use, not pretend the sample represents everyone.
Participants should understand the recording activity and the proposed use of their contributions. Voice can carry personal information, and the ability to store audio indefinitely does not remove the need for responsible consent and access arrangements. A strong resource combines technical quality with a clear, respectful relationship to the people whose speech makes it possible.
Transcription is an interpretation task
Turning speech into text requires decisions about spelling, interruptions, borrowed words and unclear passages. Different conventions can produce different transcripts from the same audio. A project therefore needs an agreed method and a way to check consistency rather than assuming that every transcriber will make identical choices.
This is especially relevant when people switch between languages. Treating every unfamiliar expression as an error can erase normal communication patterns. The team should decide how to represent those patterns in a way that serves the research question and respects the speakers’ actual language use.
Quality review can include checking a sample against the audio and discussing disagreements. The goal is not to eliminate every judgement call, which may be impossible, but to make the decisions explainable. Documentation lets later users understand what the text represents and prevents them from treating an editorial convention as a natural property of the language itself.
Text collections also carry a point of view
A text dataset reflects what was available and selected. News articles, official documents, literature and everyday messages contain different vocabulary and styles. Combining them without recording their origins can make it difficult to understand why a tool performs well on one task and poorly on another.
For illustration, a spelling tool developed mainly from formal writing may be less helpful when users type conversational messages. That does not automatically make the resource defective. It means the intended application and the training material need to be considered together.
The cultural question is equally important. A digital system should not treat the most easily collected material as the only legitimate form of expression. Researchers and developers can acknowledge the limits of a collection, seek appropriate additional material and involve people with relevant language expertise. Good resource building makes the selection visible rather than hiding it behind a claim of comprehensive coverage.
Descriptions make resources discoverable
A dataset that cannot be found or understood has limited practical value. Metadata describes what the resource contains, who created it and how it can be accessed or used. SADiLaR’s guidelines emphasize documentation, discoverability and management of digital language resources.[3]
Useful descriptions should answer a future user’s ordinary questions. Which language varieties are included? What period or setting does the material reflect? Are there known limitations? Which files belong together? The details should be specific enough to prevent accidental misuse without making the resource impossible to navigate.
Version information also matters. If a collection is corrected or expanded, researchers need to know which version underlies a result. A stable reference supports reproducibility and helps contributors receive credit. This administrative work is part of the scientific value of the resource, because it allows someone else to inspect or repeat the reasoning built on it.
Open access still requires attention to conditions
SADiLaR describes open access to its repository while requiring resource-level licensing and submission documentation.[2] Readers should not interpret “available online” as meaning that every file has the same permissions. The relevant license and access conditions should be checked before reuse.
This is particularly important when a research resource is considered for a commercial product. The intended use may differ from the purpose under which material was originally collected or shared. A developer needs to establish the applicable permissions rather than assume that a download button settles the question.
Responsible reuse also includes attribution and preserving important context. A resource can represent substantial work by speakers, annotators, linguists and technical teams. Recognizing those contributions makes the research chain more visible and helps prevent the final product from appearing to have created its linguistic foundation without any human labor behind it.
Digital participation includes creating knowledge
Language technology is not only about making people consumers of translated interfaces. It also concerns their ability to write, edit and share knowledge in their own languages. SADiLaR’s project information describes the SWiP collaboration involving SADiLaR, Wikipedia and PanSALB.[4] The initiative connects digital participation with the presence of language content online.
The broader idea is that a language becomes more useful digitally when people can create meaningful material in it, not only read material translated from elsewhere. That work involves subject knowledge, editorial judgement and community participation alongside technical access.
For a hypothetical local-history contribution, the challenge would include finding reliable sources, choosing appropriate terminology and following the platform’s editorial rules. Technology makes publication possible, but it does not automatically supply evidence or resolve disagreements. A sustainable contribution process needs people who can review, improve and maintain the material after the initial workshop or campaign ends.
Evaluation should resemble actual use
A language tool may perform well on a benchmark while disappointing users in a different setting. Evaluation should therefore include material that resembles the intended task and should explain what was measured. A transcription score, a translation judgement and successful completion of a search are different outcomes.
For example, a voice interface for a limited set of appointment requests may need a different evaluation from a general transcription service. The relevant question is whether users can complete the intended interaction reliably, including when they phrase a request differently from the examples used in development.
Testing should also look beyond an overall average. A single score can conceal differences among speakers, topics or environments. The team should investigate where errors cluster and decide what can be improved or needs a clearer limitation. Honest evaluation gives developers a better direction than a broad claim that the system “understands” a language in every context.
Human review remains valuable in important communication
Automatic output can be useful without being ready for every public-facing purpose. A draft translation may help a person begin, while important instructions still need review by someone with the relevant language and subject knowledge. The appropriate level of review depends on the consequences of an error.
Meaning can be lost through a small wording change. A public notice about dates, eligibility or required documents needs more than fluent-looking sentences. Review should check whether the message preserves the intended action and whether readers can understand it in context.
This is also why language expertise belongs early in product development. Asking for a final translation after the interface is fixed may reveal that buttons are too short, examples are culturally confusing or the information structure does not travel well. Collaboration from the beginning can improve both the language and the underlying service design.
Students can contribute resources that outlast a thesis
SADiLaR’s student repository information encourages research data to be made available with appropriate structure and licensing.[5] This can extend the usefulness of work that might otherwise remain on an individual computer after a qualification is completed. The value depends on preparation, not merely uploading a folder at the end.
A student can plan early for documentation, permissions and a clear separation between shareable material and restricted information. Keeping those needs in mind during collection makes later deposit easier. Retrospectively reconstructing consent or file relationships can be difficult once participants and collaborators are no longer readily available.
The benefit reaches the next researcher, who may be able to build on the resource instead of recreating it. It also benefits the original contributor through visibility and a clearer record of the work. Shared infrastructure is strongest when individual projects leave behind something understandable enough for others to use responsibly.
A language tool needs maintenance after launch
Language use changes, and digital services change with it. New terms, new content and revised user needs can expose gaps in a tool that once worked adequately. Maintenance should include a way to collect feedback, prioritize corrections and document changes.
An organization should also know who is responsible when an error is reported. If language support belongs to no one after launch, mistakes can persist even when users identify them clearly. A practical feedback route can turn community knowledge into improvement, provided the organization responds and does not expect unpaid users to carry the entire burden.
Sustainability includes preserving resources in usable formats and maintaining the information needed to interpret them. A dataset is not secure merely because several copies exist. Future users need compatible files, documentation and an institution prepared to manage access and preservation over time.
Inclusion is demonstrated through usable tasks
South African language technology is advancing through the less visible work of collecting, documenting, sharing and evaluating resources. SADiLaR’s role shows why infrastructure matters: many useful products depend on foundations that no single application team could responsibly invent overnight.[1][2][3]
The most meaningful outcome is a person completing a task in a language they understand and use confidently. That might involve finding information, contributing knowledge, reviewing a document or interacting with a service. A language option becomes valuable when it works through the whole task, including questions and errors.
This places cultural knowledge and technical design in the same conversation. Better systems need capable software, but they also need speakers, researchers, editors and communities whose judgement shapes what the software learns and how it is used. The advancement is a digital environment that can accommodate more of South Africa’s linguistic life without asking everyone to leave that life behind at the login screen.
Sources
[1] SADiLaR. Official mandate and research infrastructure overview. Checked 23 September 2026.
[2] SADiLaR. Resource guidelines and repository structure. Checked 23 September 2026.
[3] SADiLaR. Guidelines for digital resources, metadata and preservation. Checked 23 September 2026.
[4] SADiLaR. Activities and projects, including SWiP. Checked 23 September 2026.
[5] SADiLaR. Student data repository. Checked 23 September 2026.