What is Low-Resource Language?

Low-resource does not mean few speakers. Many low-resource languages have tens of millions of native speakers; what is scarce is digitized public corpus.

Data for these languages has to be collected from scratch, which makes it expensive, and also means there is little competition, so what you build is a scarce asset.

Degrees of low-resource

Low-resource describes the corpus, not the population. A language with tens of millions of speakers can still be low-resource if little of what is spoken is written or transcribed; a language with far fewer speakers can be well covered if it has an active transcription tradition. The label describes how much a model can learn from public material, not how many people speak the language.

There are degrees. A language may have a large written tradition but no speech corpus, which makes the job collection and alignment. At the other end, a language may have no settled writing system at all, which changes the task entirely: the orthography has to be agreed with the community before a transcription convention can even be written.

Scarcity also exists inside well-resourced languages. Medical conversations, legal proceedings and many regional dialects are low-resource even though the language around them is not: there is plenty of French in the world, and very little recorded medical French. The test is the same in every case, which is whether enough material exists to train from, or whether it has to be collected.

Where it shows up in a quote

Low-resource work is priced per project rather than per hour, because everything before the recording has to be built: recruiting speakers, writing the script, agreeing the spelling convention, training the annotators. The same volume costs a multiple of what it would in a well-resourced language, and the quote should be read in that light rather than compared hour for hour.

The speaker pool is a design decision, not a sourcing detail. For a scarce language the buyer should say which region and which speaker profile the data has to represent, because a set collected from one city may not match the users the model will meet. Dialect naming belongs in the brief: the region, not just the language name.

Ask what else the project produces. A low-resource collection usually creates assets along the way, such as a written transcription convention or a small aligned corpus, and those are worth keeping regardless of what happens to the model. What is collected well in a scarce language is a scarce asset, and it stays valuable as long as the rights allow reuse.