Resources
Low-resource language data: pricing and collection planning
Scope speech and text collection by named language, dialect, native review, accepted units and participant compensation.
Published 2026-10-27 · 6 min read
| Brief field | Specify explicitly |
|---|---|
| Language target | Named language, variety, script and code-switching |
| Source type | Read speech, prompted speech, conversation, text or parallel text |
| Review | Qualified native reviewers, orthography and adjudication |
| Price unit | Accepted speech hour, document or aligned segment; work priced separately |
Low-resource describes a gap relative to a task and available resources, not one uniform product or a small number of speakers. Name the language, regional variety, domain and model objective before asking what the data costs. Speech recognition, document understanding and translation need different deliverables.
No universal scarcity multiplier, per-language fee or collector wage is verified here. Separate the dataset licence, participant compensation, collection logistics, transcription, native-language review and delivery costs. A rare language label alone does not establish buyer demand or a sale price.
Use named language evidence rather than a broad scarcity claim
Google’s February 2026 WAXAL announcement identifies Acholi, Hausa, Luganda and Yoruba among its covered languages and describes local collection partnerships and picture-prompted speech. This is an example of a documented collection approach, not proof that those languages meet your domain or dialect needs.
Existing public material can change what needs collecting. Inspect the dataset documentation, actual language coverage and terms for the intended task before commissioning a duplicate corpus. Public availability does not establish permission to resell it or an earnings rate for new contributors.
Define language, dialect and recording style
In the brief, distinguish the language from regional or community variety, code-switching, writing system and target domain. State whether the source is read speech, prompted descriptions, spontaneous dialogue, text or paired translation. Do not combine these into one undifferentiated hour count.
Use anonymous speaker IDs to track coverage without disclosing unnecessary identities. Agree the recruitment and permission process locally, including how the proposed uses will be explained in a language participants understand. Do not collect first and assume use permission can be reconstructed later.
Budget for qualified native review
Proposed workflow: engage reviewers fluent in the target variety and task domain, use an agreed orthography guide and test a pilot containing difficult examples. Define code-switching, borrowed words, hesitations and uncertain passages rather than forcing everything into one spelling convention.
Have a second qualified reviewer inspect a sample and resolve disagreements under a documented policy. Record reviewer qualifications and review status without publishing unnecessary personal details. This is a recommended project workflow, not a claim that Fiund already has a staffed team for every language.
Price the accepted deliverable and the work behind it
A speech quote should identify elapsed versus speech-only hours, distinct speakers, transcript level, timestamps and whether translation is included. A text quote should define documents or segments, source permissions, parallel alignment and quality review. Translation is a separate deliverable from transcription.
Hypothetical cost comparison: $900 of reviewed transcription over 15 accepted speech hours is $60 per accepted hour for that work. It is neither a market benchmark nor the licence value or participant pay. Budget participant time, recruiting, recording, review and recapture as explicit lines with locally agreed terms.
Run a pilot before expanding recruitment
Review a representative pilot with target-language experts and the buyer. Include expected dialects, acoustics and domain vocabulary, then agree acceptance, correction and payment rules. Automatic scores alone may miss orthographic or regional errors.
Buyers can submit a language-specific brief with the fields below. Collection teams can describe the languages they can genuinely recruit and review. A capability submission does not prove qualified staffing, guarantee an assignment or authorize speculative recording. Fiund’s standard terms prohibit synthetic voice or likeness replication.
Before requesting an assessment or quote
- Name the target variety and domain.
- Check relevant existing resources and their terms.
- Confirm qualified native review and a pilot.
- Separate participant pay from dataset licence and review costs.
Frequently asked questions
Does low-resource always mean a higher price?
No. Task fit, source quality, permissions, coverage and available alternatives matter. No universal premium is established here.
Is translation included with transcription?
Only if the scope says so. Transcription records speech in its language; translation creates a different-language rendering.
Can Fiund collect every named language?
Availability and qualified collection or review capability must be confirmed. This article is not an inventory or staffing claim.