Speech data collection
Scripted and spontaneous Khmer recordings from speakers across provinces, ages and genders — phone, mobile and far-field.
- Custom scripts & wake words
- Conversational & call-centre audio
- Balanced demographics
Khmer AI training data
The Khmer data your AI has been missing.
High-precision Khmer data, voice and AI services: data collection, transcription, labelling, voice generation, subtitling & dubbing, and AI prompt engineering.
- Native Khmer linguists
- Every speaker consented
- Free pilot batch
bank_call_0142.wav
QA passed
-
សួស្តី ខ្ញុំចង់បើកគណនីធនាគារ។
“Hello, I'd like to open a bank account.” -
បាទ សូមបង្ហាញអត្តសញ្ញាណប័ណ្ណរបស់អ្នក។
“Sure, please show your national ID card.”
Built for teams in
- Banking & fintech
- Telecom & call centres
- Voice assistants
- Media & e-learning
- Government services
- AI research labs
Partners we've worked with
1 dayReply to every brief within one business day
FreePilot batch before you commit to anything
2×Linguist review on every transcript and label
100%Contributors consented and paid fairly
Services
Everything Khmer, from raw data to finished voice and AI
Eight services delivered by native Khmer specialists. Use one or combine them. Every project starts with a small free pilot so you can check quality first.
Transcription
Verbatim or clean-read Khmer transcripts with time-stamps, speaker turns and consistent spelling rules.
- Segment & word-level timing
- Khmer–English code-switching
- Noise & non-speech tags
Annotation & labeling
Text, audio and image labels designed with your ML team and checked by a second linguist.
- Intent, entity & sentiment
- Khmer word segmentation
- Bounding boxes & key-points
Text & LLM data
Khmer instruction data, translations and human preference ratings to make chatbots and LLMs speak natural Khmer.
- Prompt–response writing
- Human translation & review
- Output ranking (RLHF)
Evaluation & benchmarks
Held-out Khmer test sets and human scoring so you know how well your model really performs before launch.
- WER / CER test sets
- Accuracy & bias checks
- Vendor model comparisons
New
Khmer voice generation
Natural Khmer voices for apps, IVR, audiobooks and assistants: studio recordings from consented voice talent, ready-to-train text-to-speech data, and custom brand voices.
- Studio TTS datasets (48 kHz, phonetically balanced)
- Custom brand voice with signed voice-licence consent
- Khmer voiceovers for IVR, e-learning and ads
- MOS listening tests for synthetic voices
New
AI engineers & Khmer prompters
Native Khmer prompt engineers and AI engineers who design, test and harden your chatbots and AI agents so they answer correctly, politely and safely in Khmer.
- Khmer system prompts and few-shot libraries
- Red-teaming and safety testing in Khmer
- RAG and fine-tuning on your Khmer documents
- Prompters embedded in your team, by the month
New
Subtitling & dubbing
Khmer subtitles and dubbing for films, series, e-learning and ads, plus subtitles in English and Chinese for Khmer content going abroad.
- Timed subtitles (SRT, VTT) and closed captions
- Khmer dubbing and voiceover with lip-sync
- Khmer ⇄ English and Chinese translation
- Native QA on timing, spelling and tone
Ways to work with us
Pick the model that fits your team
Buy data that is ready today, hand us a whole project, or add Khmer specialists to your own pipeline.
Licence a dataset
Off-the-shelf Khmer speech, voice and text data with a datasheet and commercial licence.
Best when you need to start training this month.
Browse datasets →Managed project
We run collection, transcription, labelling and QA end to end, to your spec and deadline.
Best when you need custom data and a fixed price.
Estimate a project →Dedicated team
Reserved Khmer linguists and annotators trained on your guidelines, month to month.
Best when your data needs are ongoing.
Talk to us →Inside your tools
Our team works in your annotation platform or cloud, so sensitive data never leaves your systems.
Best when data is regulated, like banking or ID.
See security →Why specialists
Khmer breaks generic data pipelines
Global vendors treat Khmer as just another language. It isn't — and the mistakes show up as poor accuracy in production.
- ខ្ញុំស្រលាញ់កម្ពុជាNo spaces between words
- Word boundaries must be decided by trained linguists, or tokenizers and search fail.
- ស្ត្រីStacked consonants
- Subscript letters and Unicode ordering errors silently corrupt text and OCR labels.
- OK បងCode-switching
- Real speech mixes Khmer with English, French, Thai and Vietnamese words — and chats use Romanized Khmer.
- ទឹកRegional accents
- Phnom Penh, Battambang, Takeo and Khmer Krom speakers sound different. Your data needs all of them.
- កា · គាTwo consonant series
- The same vowel sign sounds different after first- and second-series consonants. Text-to-speech and pronunciation models need linguists to get it right.
- ញ៉ាំ · ឆាន់ · សោយRoyal and monk vocabulary
- "Eat" is a different word for ordinary people, monks and the king. Chatbots and subtitles must choose the right register.
- ១២៣ · 123Khmer digits and dates
- Khmer and Arabic digits, the Buddhist calendar and lunar dates all appear in real text and speech. They must be normalised the same way every time.
- Limon → UnicodeLegacy fonts and hidden characters
- Older documents use pre-Unicode Limon fonts, and web text hides zero-width spaces. Without cleaning, one word gets stored many different ways.
- ហ្វេសប៊ុកNames and loanwords
- Pali, Sanskrit, French and English loanwords, and people's names, have several accepted spellings. A project glossary keeps them consistent.
- រតនគិរីVery little open data
- Most public Khmer text is news and social media, and recorded speech from outside Phnom Penh is rare. Good models need data collected on purpose.
Use cases
What our data powers
Speech-to-text and voice assistants that understand Khmer
Train ASR and wake-word models on hours of diverse, accurately transcribed Khmer — recorded on the same phones and in the same noisy places your users live.
- Hours of transcribed read & spontaneous speech
- Wake-word & voice-command sets
- Pronunciation lexicons
Turn Khmer customer calls into insight
Transcribed and labelled call recordings let you build automatic summaries, quality scoring and complaint detection for Khmer-speaking customers.
- Speaker-separated transcripts
- Intent & sentiment tags
- PII redaction
Chatbots that write and reply like a Cambodian
Native writers create Khmer instructions, answers and ratings so your assistant sounds natural, polite and culturally correct.
- Instruction & dialogue data
- Polite-register & style guides
- Safety & quality evaluation
How we work
From brief to dataset in five steps
- 1
Scope
We agree the data spec, label guide, volume and timeline with your team.
- 2
Pilot
A small free batch so you can test quality and refine the guidelines.
- 3
Collect
Consented recording and capture with our vetted contributor network.
- 4
Label & QA
Two-pass linguist review with measured agreement on every batch.
- 5
Deliver
Secure hand-over in your format, with a quality report and data card.
Quality & security
Data you can put into production — and defend
Informed consent, every time
Contributors sign a clear Khmer-language consent form and are paid fairly. You receive the consent records.
Measured accuracy
Double review, gold-standard checks and agreement scores reported per batch — not just promised.
Locked-down handling
Encrypted storage and transfer, role-based access, NDAs for all staff and deletion on request.
Data stays where you need it
Work on our secure platform or inside your own tools and cloud.
What you receive with every batch
Quality you can measure, not just trust
batch_07_report.pdf
Accepted
Transcript accuracy98.7%Target ≥ 98%
Annotator agreement (κ)0.91Target ≥ 0.85
Gold-task pass rate99.2%Hidden checks in every batch
Consent records412 / 412One per speaker, delivered to you
Illustrative example. Your report shows the real numbers for each batch, with error breakdowns by speaker, region and label.
Project estimator
Get a ballpark in 30 seconds
Pick a service and volume to see an indicative budget and delivery time. You get a fixed quote after the free pilot.
Indicative budget
—
Estimated delivery
— weeks
- Free pilot batch first
- Datasheet, consent records and QA report
- Delivery in your format
Indicative only. Rates depend on audio quality, label complexity and turnaround.
FAQ
Common questions
How does the free pilot work?
We produce a small sample batch to your spec — for example 1 hour of transcribed audio or 500 labelled images — at no cost. You review it, we refine the guidelines, then we quote the full project.
How do you price projects?
Speech is usually priced per audio hour, labelling per item, and collection per contributor or per hour recorded. Volume, turnaround and difficulty affect the rate. You get a fixed quote after the pilot.
Can you work with our own annotation tool?
Yes. Our team can work inside your platform, or we use our own secure tools and deliver in your format.
Who owns the data?
For custom projects you own the delivered dataset, with consent records that allow the agreed use. We don't resell custom data.
Do you cover other languages?
Khmer is our focus. We can also help with Khmer–English parallel data and minority languages spoken in Cambodia on request.
Start a project
Tell us what you're building
We'll reply within one business day with questions and a pilot plan.
- Email hello@example.com
- Telegram @example
- Based in Phnom Penh, Cambodia