Buying data for low-resource languages: what changes

Low-resource does not mean few speakers. Navajo has fewer than 200,000 speakers, but so do several languages with well-developed digital corpora. What makes a language low-resource is the absence of existing text and audio that a model can learn from.

For a buyer, that changes the shape of the project more than any other single factor.

The bottleneck moves

In a well-resourced language, the constraint is usually studio time and annotation throughput. Speakers are easy to find and the recruitment step is a formality.

In a low-resource language the constraint is almost always recruitment. Qualified native speakers may be geographically concentrated, may not be reachable through standard channels, and may have limited availability. Finding forty of them can take longer than recording all of them.

This is why timelines for low-resource projects are unreliable in a specific way: the estimate is dominated by a step whose duration is hard to predict in advance.

Existing text is not a shortcut

When a language has some written material (religious texts, colonial-era documentation, community publications), there is a temptation to use it as the transcription target.

It usually does not work. Older written material tends to be in a register nobody speaks, in an orthography that has since changed, or in a dialect other than the one you are recording. The audio and the text end up describing different things.

What to expect on quality

The realistic quality target for a first collection in a low-resource language is lower than for a well-resourced one, and that is a property of the situation rather than of the producer.

• Annotator agreement will be lower, because the orthography may not be fully standardized and there is no established convention to fall back on.
• Some of the annotation budget will go to establishing conventions rather than applying them.
• A pilot batch matters more here than anywhere else, because it is where the conventions get tested.

The offsetting advantage

Low-resource data is scarce, which means it is also defensible. A competitor cannot buy the same corpus from a marketplace, because there is no marketplace listing for it.

For teams building a capability that depends on language coverage, that scarcity is the point. The project is slower and more expensive per hour, and the resulting asset is one that cannot be replicated by writing a larger purchase order.

Where the speakers actually are

Recruitment in a low-resource language rarely works through the channels that work elsewhere. Job boards and participant panels are thin or absent, and the speakers may be geographically concentrated in ways that a national campaign does not reach.

The channels that do work, in rough order of yield: diaspora organizations in cities with an established community, university linguistics and language departments, community and religious institutions, and existing speaker networks reached through one trusted introducer. The last one is the most effective and the least controllable, because it depends on a person rather than a process.

Two things belong in the plan from the start rather than being added later. First, compensation at a rate that reflects the scarcity of the skill rather than the local minimum wage; underpaying a small pool is how a project loses access to it permanently. Second, a plain statement to the community about what the recordings will be used for and who will hold them. In small speaker populations, word travels, and a project that is perceived as extractive will not get a second round of participants.

Structuring the project around the recruitment step

Because recruitment dominates the timeline and recruitment duration is not predictable in advance, a fixed delivery date is the wrong contract shape for a first collection. It converts an uncertainty into a penalty, and the penalty lands on whichever party has less leverage.

The structure that holds up instead is a sequence of gated phases, each with its own exit condition: a recruitment pilot that stops once a target number of screened speakers have completed a session, a convention-setting batch that produces the annotation guideline and tests it against real audio, and a main run that proceeds only after the guideline is signed off. Payment attaches to each gate rather than to the calendar.

This costs the buyer a little schedule certainty and buys something more useful: an early, cheap signal about whether the project is viable at all. If the recruitment pilot cannot find twenty qualified speakers in three weeks, that is a finding worth having before the annotation budget is committed, not after.