AI Training Data Providers

The five questions that separate providers

Most provider websites look alike. The differences show up in the answers to five questions, and any provider worth using will answer all of them in writing.

• How many distinct speakers are in the dataset? Hours are easy to inflate by recording the same people longer. Speaker count is the number that determines whether a model generalizes.
• Who are the annotators, and are they native speakers of the target language? A fluent second-language speaker and a native speaker produce measurably different transcription quality.
• What is the annotation guideline, and can I see it before the project starts? If they will not show you the guideline, they either do not have one or it is not good enough to share.
• What happens if the delivery does not meet spec? A provider with a real quality process will describe a re-work policy. One without will change the subject.
• Can you show the consent and licensing chain? For any data involving real people, this is the difference between a usable asset and a legal liability.

How we answer those five questions

These are the same five questions, answered the way we would want them answered if we were the ones buying.

Where we sit in that market

We are an AI data and trading platform. We run our own collection and processing, and we also work with vetted producers, which means we are not defending a fixed catalog or a fixed set of recording studios. The project gets matched to the capability that can actually meet it, wherever that capability sits.

It also means we can be honest about coverage. For the 31 languages with documented search demand we have deep production paths and can usually quote within days. For the 89 coverage languages, we can often produce but the timeline is longer and we will say so.

What a serious specification looks like

Buyers who get good results tend to write specifications that answer the same questions we ask. Here is the shape of one.

• Language, dialect, and the region speakers should come from.
• Total hours, and minimum distinct speaker count.
• Recording conditions: studio, home, mobile, or a specified noise environment with a target SNR.
• Annotation type: transcription only, transcription plus intent, phoneme-level, or diarization.
• Delivery format and sampling rate, plus whether you need speaker metadata per file.

Red flags on the buyer side too

Providers are wary of the same things buyers are. A request that asks for a price before specifying hours and annotation depth will get a slow, hedged answer from anyone experienced, because there is no honest number to give.

Requests that require the provider to figure out the use case after delivery, or that leave the compliance chain undefined, tend to stall for the same reason.

What changes when you switch providers

Buyers rarely change provider on price alone. The cost of switching is usually higher than the saving, and it is concentrated in the first project.

Understanding where that cost sits tells you whether a switch is worth making, and what to ask a new provider to demonstrate before committing.

• A new specification round. A provider who has worked with you knows your format and your quality bar. A new one starts from a blank document.
• A pilot that cannot be skipped. The first batch with a new provider has to be reviewed closely, which takes your team's time as well as theirs.
• Documentation that has to be re-checked. Consent records and metadata formats differ between providers, and the differences surface at the worst moment.
• Delivery format mismatches. Small differences in file structure or metadata naming break downstream pipelines quietly.
• The risk of a second source. Running two providers in parallel is the usual way to reduce dependence, but it doubles the review work.

What to ask before committing to a new provider

These four questions cover most of what decides whether a switch goes smoothly. They are also the questions a provider who has done this before will already have answers to.

A provider who cannot answer them is not necessarily a bad provider. They are simply one you will have to manage more closely during the first project.

• Can I have a sample in my format, not yours? The answer shows whether they have worked with buyers whose pipelines are already built.
• What does your metadata schema contain? Ask before the first project, not after. Retrofitting metadata is expensive.
• Who holds the consent records, and do I receive copies? You should hold them. Holding them behind the provider creates a dependency you cannot audit.
• How is the first project structured? A defined pilot with a review point is the sign of a provider who expects the first project to teach them something.