Pricing · speech, text, image, annotation
How AI training data is priced, and what ours includes
There is no price list, and any vendor who publishes one is hiding the variables. This page shows exactly how a quote is built, what every delivery includes, what costs extra, and how to get a comparable quote from us in two business days.
Three ways to start
Free sample, paid pilot, production programme
Every engagement starts with the sample. Most move to a pilot within two weeks. Pilot fees are credited against production.
Free sample pack
No cost, no watermark, no sales call
- 2–5 hours of audio, 1–2k text items or ~500 images
- Full metadata, transcripts and datasheet
- Consent form and licence text included
- Delivered within one business day
- Identical format and quality to the paid corpus
Paid pilot
Quoted per hour, per thousand items or per image
- 50–200 hrs speech · 25k–100k text items · 10k–50k images
- One or two languages, 3–5 weeks
- Spec, quota and acceptance test signed before recording
- First QA-passed batch in week 2–3
- 100% credited against a production order
Production programme
Tiered per-unit pricing, milestone-billed
- 1,000+ hours, 500k+ items or 100k+ images
- Dedicated programme manager and shared channel
- Weekly batches with QA report and running datasheet
- Optional exclusivity, on-prem or in-VPC work
- Multi-year framework agreements available
Units
What each data type is priced by
Quotes are always per QA-passed unit. Rejected units are not invoiced.
| Data type | Priced per | What moves the rate | Always included |
|---|---|---|---|
| Speech for ASR | Hour of accepted audio | Language rarity; spontaneous vs read; quota strictness; diarization; second transcription pass; telephony vs wideband | Recruitment, recording, verbatim transcript, metadata, consent, QA |
| Speech for TTS | Hour of studio audio | Number of voices; script design; 44.1/48 kHz studio time; pronunciation lexicon | Voice casting, script, recording, alignment, lexicon |
| Text corpora | Thousand words | Domain; native authoring vs sourcing; register; de-duplication and PII pass | Provenance, licence, PII redaction, QA |
| SFT / instruction data | Prompt–response pair | Turns per dialogue; domain expertise; length; rubric complexity | Native authoring, dual review, metadata |
| RLHF / preference data | Ranked comparison | Number of responses ranked; rater qualification; agreement floor | Rater training, agreement stats, rationale text |
| Parallel / translation | Aligned sentence pair | Language pair; domain; human vs post-edited; back-translation QA | Alignment, review, domain tags |
| Image & video | Image, or minute of video | Scenes and conditions; consent complexity (faces); annotation depth per frame | Capture, consent, EXIF/metadata, QA |
| OCR / documents | Page | Printed vs handwritten; line vs word vs character ground truth; document sourcing | Scanning, transcription, layout boxes |
| Annotation on your data | Item, or annotator-hour | Task complexity; qualification; passes; tooling; security tier (remote, VPC, on-prem) | Guidelines, calibration, agreement reporting |
Variables
The five things that move a quote
Language and variety
Urdu in Karachi is the cheapest thing we field. Brahui in Kalat or Thari Sindhi requires bespoke recruitment and costs several times more per hour.
Register and task
Read speech from prompts is fast. Spontaneous two-party conversation, or native-authored multi-turn SFT, takes more contributor time and more QA.
Quota strictness
A 50/50 gender split nationwide is easy. Equal cells across four age bands, urban/rural, three dialects and two device classes means rejecting clean audio that lands in a full cell.
Annotation depth
Verbatim transcript alone, or transcript plus diarization, noise tags, code-switch spans and a second independent pass. Each layer adds skilled hours.
Security tier and speed
Remote annotation is standard. In-VPC, on-prem or secure-facility work, and rush timelines that need parallel teams, carry a premium.
Read the full buyer's guide
The costs that never appear in a quote, and an RFQ template that gets comparable bids.
What AI training data actually costsIncluded in every quote
- Per-record consent artefacts, versioned and auditable
- Datasheet: composition, method, QA results, known limitations
- QA report against the signed spec
- 14-day acceptance window; out-of-tolerance units re-collected free
- Perpetual, worldwide, sublicensable training licence with derivative-model rights
- Delivery to your S3, GCS or Azure bucket with manifests and checksums
- DPA, and SCCs or UK Addendum where required
- PII redaction pass on text and transcripts
Priced separately, stated up front
- Exclusivity, by period or field of use
- Secure-facility, on-prem or in-VPC collection and annotation
- Rush delivery requiring parallel field teams
- Third and further transcription or review passes
- Specialist domains needing qualified raters (medical, legal, finance)
- Studio TTS voices with professional voice talent
- Custom tooling or integration with your annotation platform
- Long-term hosting or re-delivery after the retention period
Get a quote
What to send for a quote you can compare
Two lines is enough to start. The seven points below get you a firm per-unit price in two business days.
Language and variety
Urdu; Punjabi Majhi plus Pothwari; Sindhi Vicholi and Lari; and so on.
Volume and modality
Hours of audio, thousands of text items, images or pages.
Register and conditions
Spontaneous or read; telephony or wideband; noise and device mix.
Quota axes
Age, gender, urban/rural, region, dialect, device — and how strict.
Annotation depth
Transcript layers, diarization, tags, number of passes, agreement floor.
Acceptance test
How you will judge the delivery: held-out set, WER target, agreement threshold.
Timeline and security tier
When you need the first batch and the last, and whether the work must happen inside your perimeter.
Answers
Questions buyers ask about price
Because the same hour of audio can vary several-fold in cost depending on language rarity, register, quota strictness, transcription depth and QA level. A published number would either mislead you or be padded to cover the worst case. We quote against a written spec instead, within two business days.
Yes. Every catalog dataset has a free, unwatermarked sample pack — typically 2–5 hours of audio, 1–2k text items or around 500 images with full metadata — delivered within one business day with no sales call. It is identical in format and quality to the paid corpus.
A paid pilot: usually 50–200 hours of speech, 25k–100k text items or 10k–50k images in one or two languages, running 3–5 weeks. Pilot fees are credited in full against a production programme.
Per hour of QA-passed audio, inclusive of recruitment, recording, verbatim transcription, metadata, consent artefacts and QA. Rarer languages, spontaneous conversation, strict demographic quotas, diarization and second transcription passes each move the rate up; read speech in Urdu with a single pass sits at the lower end.
Per item: per thousand words for corpora, per prompt–response pair for SFT, per ranked comparison for RLHF, per aligned sentence pair for parallel text. Native authoring costs more than translation and is worth it; see the SFT data quality guide.
Pilots are invoiced on delivery. Programmes are milestone-billed against accepted batches, typically net 30 by international wire, in USD by default or GBP, EUR and AUD on request. Rejected units are never invoiced.
Yes. Standard licences are non-exclusive, perpetual, worldwide and sublicensable for training. Exclusivity for a period or a field of use is available at a premium and is the only thing that changes the licence rather than the delivery.
Send every vendor the same spec: language and variety, hours or items, register, quota axes, acoustic conditions, transcription depth, acceptance test and timeline. Ask each to price the same unit and to state what is excluded. Our buyer's guide to training data cost includes an RFQ template.
Get a per-unit quote in two business days
Send the spec, or just the language and the problem. We come back with a firm price per QA-passed unit, a timeline, and a free sample so you can check the quality before you sign anything.
Pilot fees are credited in full against production. Rejected units are never invoiced.