Written for vendor and supply managers at multilingual vendors and AI data companies. What we have delivered as a subcontractor, how a batch runs inside your workflow, and what we can return when you send an onboarding pack.
Our AI data production to date has run as a company-to-company subcontract under a global language-services provider, feeding major AI programs. We do not name the programs or the parties. This is the work itself.
Rolling weekly batches across seven locales — Mandarin (Mainland), Taiwan Mandarin, Cantonese, Korean, Japanese, Tagalog and Turkish, each including code-switched English. Billed by effort hour, with per-file logging.
Cantonese and Cantonese–English recording in Hong Kong, including hundreds of two-speaker service-call scripts recorded on separated tracks.
Proofreading and quality scoring of Korean–Chinese song-lyric data against a client rubric.
Ongoing on-device testing in Hong Kong with local testers, reported month by month.
A multi-month program in seven Asian language directions, each batch on a short deadline, with our internal deadline six hours ahead of the client’s to leave room for review. We recruited on demonstrated, everyday code-switching rather than declared native language, and by the second month time per task had fallen below half of the first month’s in most languages.
On the two-speaker Cantonese–English recordings, the first batch came back for headphone bleed and unsynchronised tracks. The fix went into the recording specification (a counted sync, in-ear monitoring with a level rule, a merged-track check before delivery), and later batches passed first time.
We work in the client platform when the program requires it, or on our own task platform with per-file effort logging and an export that matches your schema.
Your specification becomes a locale guide covering orthography, verbatim policy, code-switch tagging, written-form numbers, segmentation and PII handling — amended after every correction cycle so the same error does not come back.
Error categories with pass gates and sampling. For Traditional Chinese, every file passes an automated Simplified-character check before a reviewer sees it; word timestamps are checked against the waveform when the spec asks for them.
Audio and reference text are aligned before work starts. Re-cut files or revised scripts are flagged back to you rather than found after hours of validation.
On a short-deadline job we answer with a yes or no first and ask clarifying questions after accepting. We learned that a list of questions before a commitment is how a job goes to someone else.
The same local team and team lead stay on a locale from batch to batch, so context builds up instead of resetting — that is where the effort-per-unit drop came from.
Contracting entity Translia Limited (Hong Kong), with operations in Beijing through 北京传为文化有限公司. ISO 17100 and ISO 18587 certified.
We review and sign your NDA and MSA on your paper, and work through your e-signature and supplier-portal steps.
We complete capability and capacity forms per locale, mark each figure as firm or an estimate, and leave a row blank when we cannot field people for it. See our Asian-language capacity by locale.
Participant consent and provenance records per recording, and (under your NDA) a sample locale guide with a scored transcription sample.
Pricing is per effort hour for validation and per accepted unit for collection. If your target rate sits below what the speaker screening costs, we tell you that and why rather than accept and miss.
Yes. All of our AI data production to date has been company-to-company work under a larger vendor's program. We report to you, do not contact the end client, and do not name the programs we have worked on.
Seven locales have delivered production batches: Mandarin (Mainland), Taiwan Mandarin, Cantonese (Hong Kong), Korean, Japanese, Tagalog and Turkish, each with English code-switching. Seven more are ready to activate with a pilot. The capacity page lists what each locale has delivered.
No. We have no off-the-shelf dataset to license. We record, transcribe and validate to your specification, starting with a pilot batch.
A Cantonese–English recording line went from pilot to hundreds of scripts within three weeks. A new line starts with a pilot batch on your specification; volume follows once the pilot passes your review.
Validation and transcription by effort hour with per-file logging; collection per accepted unit. Invoicing follows your supplier terms.
Send the onboarding link, capability form or a pilot spec — we'll return it locale by locale.
Contact our data team →