How Khmer data work is priced
Most vendors use one of a few units:
- Per audio hour (or per audio minute) for transcription, measured on the length of the recording, not the time it takes to transcribe.
- Per recorded hour for speech data collection.
- Per 1,000 items for annotation, where an item might be a sentence, a document, an image or an intent label.
- Per prompt pair or per comparison for LLM training data.
- Per finished hour or per video minute for voice, subtitling and dubbing.
- Per day for engineers and prompt specialists working inside your team.
Always check what a unit includes. A low price per audio hour that excludes review, timestamps or speaker labels can end up costing more once you pay to fix the output.
Indicative 2026 price ranges
These are the indicative ranges shown in the estimator on our home page, in US dollars. They are a planning guide, not a quote: the final price depends on your audio, specification and volume, and is fixed after a free pilot.
| Service | Unit | Indicative range |
|---|---|---|
| Khmer transcription | per audio hour | $55–90 (about $0.92–1.50 per minute) |
| Speech data collection | per recorded hour | $40–85 |
| Annotation and labelling | per 1,000 items | $35–120 |
| Text and LLM data | per 1,000 written prompt pairs | $3,000–7,500 |
| Voice generation | per finished studio hour | $250–600 |
| AI engineers and Khmer prompters | per engineer-day | $120–280 |
| Subtitling | per video minute | $3.50–9 |
| Dubbing | per video minute | $20–45 |
For comparison, published rates for human transcription from global vendors commonly sit around $1–2 per audio minute, and public quotes we found for Khmer fall in a similar range or above. Specialist low-resource languages are often priced at a premium over English.
What drives the price of Khmer transcription
Audio quality
Clear, single-speaker audio recorded close to the microphone is the cheapest to transcribe. Phone calls, background noise, music, far-field microphones and overlapping speakers all slow transcribers down and need more review.
Speaking style and accent
Read speech is faster to transcribe than spontaneous conversation full of hesitations, slang and code-switching. Strong regional accents and specialist vocabulary (medical, legal, technical) need more experienced transcribers.
Output specification
Verbatim transcripts with fillers and noise tags, word segmentation, speaker labels, timestamps at the sentence or word level, and translation into English all add work. Decide what you really need before asking for quotes, and write it in a style guide (see how to build a Khmer ASR dataset).
Quality checks
A single pass is cheaper but leaves more errors. Double review with gold-standard checks costs more and gives you measurable accuracy. In our estimator, a 3-pass review with a gold audit adds about 20% to the price.
Turnaround
Rushed deadlines mean more people working in parallel and more coordination. In our estimator, priority turnaround adds about 30%.
What drives the price of data labelling
- Task complexity. Choosing one label for a sentence is quick; marking every named entity, segmenting every word or writing a rationale takes much longer.
- Item length. A tweet and a ten-page contract are both "one item".
- Expertise. Some tasks need linguists or domain specialists rather than general annotators.
- Guidelines. Clear guidelines with examples reduce rework. Vague ones push the cost up over the project.
- Volume. Larger projects spread set-up costs (guidelines, training, tooling) over more items.
That is why annotation ranges are wide: $35 per 1,000 simple classification items and $120 per 1,000 entity-labelled items can both be fair prices.
LLM data, voice and subtitles
Human-written Khmer prompt and response pairs cost more per item than labels because each one is written from scratch by a native speaker and reviewed. Preference ratings, where people compare model answers, are cheaper per item and are usually quoted separately. Voice data is priced per finished hour, which includes several hours of studio time, direction and editing for every hour you receive. Subtitles are cheaper when the source is already Khmer; translated subtitles and dubbing cost more.
How to budget a Khmer data project
- Define the deliverable. The model you are training, the volume (hours, items or minutes) and the output format.
- Pick the quality level. Decide what accuracy you need and how you will measure it.
- Get a ballpark. Use the project estimator to see an indicative budget and delivery time.
- Run a pilot. A small batch on your own data shows the real speed and quality. Ours is free.
- Fix the quote. Agree a price per unit, the acceptance criteria and the delivery schedule from what the pilot showed.
Leave room in the budget for a second round. Most teams find, after training their first model, that they need more of a particular accent, domain or edge case.
Cheap data is expensive later
The cheapest offer is rarely the cheapest outcome. Inconsistent transcripts, missing consent records or mislabelled entities cost engineering time, retraining and sometimes legal risk. When comparing vendors, ask for a sample on your own data, the style guide they will follow, how they measure accuracy, and how they handle consent and data security.
Get an estimate
Use the estimator for a quick range, or tell us about your project for a free pilot and a fixed quote. You can also start from our ready-made Khmer datasets, which are often faster and cheaper than collecting from scratch.