AI data services for languages your model cannot fake.

MoniSa builds datasets to a specification rather than selling from a shelf. That matters commercially: a commissioned set is collected against your task definition, your locale list and your acceptance criteria, so it does not carry the licensing history, demographic skew or consent ambiguity that a resold corpus often does. If you are shopping for existing inventory, say so early — it is a different purchase with a different timeline, and knowing which one you need changes the whole conversation.

The material that makes it usable six months later: language and dialect labels at the locale level rather than the language level, speaker demographics where the collection design captured them, the technical specification the audio or text was produced to, per-item quality status from the QA pass, and the consent and rights record covering the intended use. A dataset without its metadata is a dataset you cannot filter, re-balance or defend.

Ownership and usage rights are agreed in the contract before collection starts, not negotiated afterwards. The distinction that matters is between the right to use data for a stated purpose and the right to license it onward, because they are priced and consented differently — a contributor consenting to model training is not automatically consenting to redistribution. Tell us the downstream uses you need covered, including any sublicensing, and the consent design accounts for them from the first recording.

By deciding at specification time whether identifying content should never be captured, or captured and then redacted, because those are different collection designs and the second is more expensive to get right. Where redaction applies it is treated as a QA-checked deliverable rather than a best effort, and clinical or regulated content raises the review standard rather than just the price. If your use requires a fully de-identified set, that requirement belongs in the brief, not in a later review cycle.

Yes — demographic and locale quotas are a normal part of a collection design. Buyers routinely specify native-speaker-only, a stated regional variety, an even gender split, a spread across defined age bands, minimum unique speakers, and caps on how much any one contributor may supply so a handful of voices cannot dominate the set. Those constraints shape recruitment and timeline, so stating them up front produces a realistic schedule rather than an optimistic one.

Whatever the receiving pipeline requires, stated as a specification rather than assumed: sample rate captured natively rather than down-sampled from something else, channel configuration including dual-channel where speaker separation matters, container format, clip length bounds, and a defined noise floor for the recording environment. Synthetic or AI-generated voices are excluded unless a brief explicitly asks for them — for most training and evaluation work they are precisely what the buyer is trying to avoid.

A pilot is the normal way in, and it is worth more than a sample file. A pilot batch runs the real workflow at small scale, which surfaces the things that actually derail collection programmes: guideline ambiguity, a locale that recruits slower than expected, an acceptance criterion that reads clearly and scores inconsistently. Fixing those on a pilot is cheap. Discovering them at full volume is not.

You are told before a date is agreed, not after. Coverage is reported pair by pair as staffed today or needing a recruitment window, with the window stated — in writing, while the scope is still being agreed. Nobody new goes onto live work until a pilot batch has been reviewed and signed off. A coverage claim you cannot check before signing is not coverage.