Speech collection, transcription validation, and evaluation in Asian languages and code-switching, taken on request.
Language pair, scenario, format, and what a good transcript looks like to you. A page is enough to start.
Natural bilinguals in your register, checked at the switch points and against your transcription convention before it reaches you.
If the batch passes your review, the production line runs on the same spec. If it does not, you have spent nothing.
Scripted and spontaneous speech, dual-speaker conversational, and dialectal recordings. Managed speaker sourcing with strict technical specifications (sample rate, channel configuration, recording environment, and speaker demographics), validated per batch.
Seven code-switching locales in production today and seven more ready to activate. Natural code-switching is hard to source through general crowd platforms.
Multi-language transcription and validation QA on batch deadlines and client-defined guidelines, checked before the audio goes into training.
Adequacy, fluency, ranking, and LQA by native evaluators, scored the same way across languages.
Batch import, audio-reference alignment check, scope confirmation — mismatches flagged back before work starts.
A fixed language team claims tasks on our task platform, so context carries over from batch to batch.
Guideline-driven work with per-file effort logging, so capacity is forecastable and problem files surface early.
Second-pass review against a written variant guide that is amended after every correction cycle.
Export with an effort report and a log of corrections, so the same issue does not come back.
Locale by locale, what each has delivered: Asian-language capacity. Onboarding us as a supplier? Notes for vendor managers.
Speech recognition still struggles with code-switched speech, and natural code-switchers are hard to find by filtering a crowd on native language. Translia runs native-speaker teams across seven code-switching locales in production today, with seven more ready to activate, each with documented consent and provenance per contributor.
Cantonese–English (Hong Kong) · Mandarin–English (Mainland) · Taiwan Mandarin–English · Japanese–English · Korean–English · Tagalog–English (Taglish) · Turkish–English
Hindi–English (Hinglish) · Malay–English · Tamil–English · Indonesian–English · Thai–English · Vietnamese–English · Gulf Arabic–English
One accountable company. Vetted contributors, strict spec compliance, and per-batch quality confirmation.
Hong Kong Cantonese, Taiwan Mandarin, Simplified Mandarin, and regional variants — plus Korean, Japanese, Filipino, Turkish and a growing set. These are our core languages.
Fixed language teams give a stable core so context accumulates batch over batch — while a 10,000-linguist network absorbs weekly volume spikes without restarting onboarding. A self-serve platform handles dispatch, delivery, and QA tracking.
Documented consent per contributor and provenance per batch, so data origin and licensing can be audited.
ISO 17100 and ISO 18587 certified, with structured review on every delivery.
We support major AI platform providers and larger data companies as a subcontracted production partner, company to company.
Contact-centre voice agents across Hong Kong, Taiwan, Singapore, Malaysia and the Philippines take calls that open in one language and finish in another. Published code-switching benchmarks report word error rates several times the monolingual baseline, and the errors cluster at the switch points (the product name, the number, the borrowed word in the middle of a sentence), which is exactly where the value of a call sits. No configuration flag fixes a distribution the model never saw.
An evaluation set built from your own call scenarios, scored at the switch points rather than sentence-average, shows what the agent actually gets wrong before you commit to an engine or a fine-tune.
Adaptation data recorded by people who mix the two languages every day, in your register (support calls, appointment lines, order confirmations), speaking without a script.
Every contributor’s consent covers commercial AI-training use, with provenance per batch, so the data can go into a production model without a documentation gap.
Seven locales in production today (Cantonese, Mandarin, Taiwan Mandarin, Japanese, Korean, Tagalog and Turkish, each paired with English) and seven more ready to activate. Start with a pilot batch scoped to your call scenario.
Buyers sourcing speech data usually arrive with an hourly benchmark from general collection vendors. That number prices recording time, and it assumes contributors can be found by filtering a pool on declared native language.
Screening on “native speaker of Cantonese” finds people who speak Cantonese, but not necessarily people who switch between Cantonese and English the way your callers do. They read the script and pronounce the English words carefully and in full. The result is monolingual audio with loanwords, which general speech recognition already handles. Where the switch lands, how dense it is and what happens to the syllable before it are missing, and they cannot be added in post.
We recruit against how people actually speak day to day, in the register and scenario you need. That selection step is the expensive part of the line, and it is the part that decides whether the audio is usable at all. Recording is the cheap half.
Across a multi-month program of dozens of batches in seven Asian language directions for a major global AI program, each batch on a short deadline, with our internal deadline six hours ahead of the client’s to leave room for review. Selection was based on demonstrated switching, and by the second month time per task had fallen below half of the first month’s in most languages.
That is the argument for the price: an hour of recording is cheap, and an hour of the wrong speaker costs the whole batch.
Scaled a Cantonese-English code-switching recording program from pilot to hundreds of scripts within three weeks for a major global AI program — batches accepted with quality confirmed.
Operating a rolling multi-language transcription validation line across dozens of language variants, delivering weekly batches into a major AI platform provider's data supply chain.
Hundreds of two-speaker Cantonese-English service-call scripts, recorded locally on separated tracks. The first batch came back with rework: headphone bleed had put each speaker into the other's channel, and the two tracks did not start together. The fix went into the specification: a counted sync at the top of every take, in-ear monitoring with a level rule, and a merged-track check before delivery. Later batches passed first time.
Client programs are confidential. These describe the work, not the parties involved.
AI regulation is turning data documentation into a procurement requirement. Under the EU AI Act's transparency obligations, providers of AI systems increasingly need to show where their training data came from and on what terms. Every contributor gives documented consent, every batch has its own records, and delivery metadata is ready for audit.
Translia delivers Asian-language speech and code-switching datasets to client specification: 16–48 kHz sampling in WAV or FLAC, mono or separated dual-channel, scripted through conversational registers, with speaker composition balanced by gender, age band, and regional accent. Every batch ships with per-batch technical QA, delivery-ready metadata, and documented speaker consent and provenance records suitable for AI-regulation documentation requirements.
16–48 kHz WAV/FLAC · mono or separated dual-channel · scripted, semi-scripted, or conversational · clean and noisy environment mixes to spec.
Native speakers balanced by gender and age band · regional pronunciation controlled · code-switching pairs including Cantonese–English and Mandarin–English.
Per-batch technical QA · speaker IDs, batch IDs, environment and channel tags · transcription validation available as a second layer.
Asian languages and their regional variants (Hong Kong Cantonese, Taiwan Mandarin, Simplified Mandarin, and other Chinese variants) alongside Korean, Japanese, Filipino, Turkish, and a growing set. We also handle code-switching such as Cantonese-English and Mandarin-English.
No — there is no off-the-shelf catalogue here, and for code-switched Asian speech that is the honest answer. What decides whether a dataset solves your problem is its composition: which language pair, which register, which scenario, and how recently it was recorded. Licensable code-switching corpora concentrate heavily on a few language pairs, and much of what exists is read or synthesised speech rather than people actually switching in conversation — useful for measuring a system, rarely sufficient for adapting one (we went through the public corpora and their sources here). We build to specification instead, starting with a pilot batch scoped to your scenario. If an evaluation set is what you need and a public corpus covers your pair, we will say so.
Every contributor works under documented consent, with provenance tracked per contributor and per batch. As an ISO 17100 and ISO 18587 certified company with fixed teams, data origin, licensing, and processing are auditable.
Yes. Speech collection follows strict specs (sample rate, channel configuration, recording environment, speaker demographics, and script design) validated per batch before delivery. Transcription and evaluation follow client-defined guidelines, with QA on every batch.
Yes. We support major AI platform providers and larger data companies as a company-to-company engagement, delivering fixed-team capacity in Asian languages and code-switching audio that general vendors cannot easily source.
We run fixed native-speaker teams with one accountable company: vetted native speakers, strict spec compliance, documented consent and provenance, and per-batch quality confirmation. That matters most for the hard cases: code-switching and low-resource Asian variants.
Every batch goes through an intake alignment check. Mismatches (re-cut files, revised scripts) are flagged back to you before production starts.
Speech and validation work is typically billed by effort hours with per-file logging, so you can see exactly where time goes; collection is quoted per deliverable unit. Fixed teams give a stable core, and a wider linguist network absorbs weekly spikes without restarting onboarding each time.
Yes. Every contributor provides documented consent; sourcing and batch records are maintained for each delivery; metadata is audit-ready. We don't provide legal advice, but our production is designed so AI providers can meet transparency and provenance documentation obligations.
The commissioning client, under the licensing structure agreed per project. Speaker consent is collected in writing and covers commercial AI-training use; we retain no reuse rights unless explicitly agreed. Rights documentation (consent scope, permitted uses, and provenance records) ships with each delivery.
Audio as WAV or FLAC at the agreed sampling rate with channel separation preserved; transcripts and annotations in JSON, TSV, or your schema; per-item metadata including speaker ID, demographic band, environment and channel tags, plus batch-level QA results. Batches are delivered by secure transfer on the cadence your pipeline needs.
Only what you commission us to produce is training data — never your own material. Any client content shared with us for reference, specification, or evaluation work is processed under standard confidentiality terms and, where any AI tooling is involved, only through enterprise API tiers configured with zero data retention (ZDR) and model-training opt-out; we never put client material through free or consumer AI tools. Datasets we deliver are licensed to the commissioning client, with consent scope and provenance records included.
For sensitive or regulated work we can run a fully human-only workflow with no AI in the chain. On request we disclose the specific engines used under NDA and sign your own Data Processing Agreement (DPA).
Yes, this is a core line. Natural code-switchers are hard to find by filtering a crowd on native language, so we use native-speaker teams briefed on scenario and register. We run seven code-switching locales in production today (Cantonese–English, Mandarin–English, Taiwan Mandarin–English, Japanese–English, Korean–English, Tagalog–English, Turkish–English) with seven more ready to activate: Hindi–English, Malay–English, Tamil–English, Indonesian–English, Thai–English, Vietnamese–English, and Gulf Arabic–English.
Scripted, spontaneous, and dual-speaker conversational formats are all supported, with documented consent and provenance per contributor. We can start with a free pilot batch so you can check switch-point handling and transcription convention against your spec before committing.
Treat it as a data problem. General-purpose speech recognition degrades exactly at language-switch points (embedded product names, numbers, borrowed words) because it is trained on monolingual corpora, and no configuration flag fixes a distribution the model never saw. The workable path is to benchmark on real code-switched audio from your own call scenarios, measured at the switch points rather than sentence-average word error rate, and then adapt the model with training data recorded by natural bilinguals to your scenario and register.
That recording-to-spec production is what we do: appointment lines, customer support calls, and conversational formats, briefed to your domain vocabulary, across the code-switching locales listed above.
Three tests, and none of them require naming brands. First, benchmark on real code-switched audio from your target locale (not the vendor's demo set) and score accuracy at the language-switch points, not sentence-average. Second, check whether named entities and borrowed words survive: product names embedded in the other language are where the commercial value of a call concentrates, and where models tend to "correct" them into noise. Third, ask what data the system can be adapted with — if a vendor cannot fine-tune on natural code-switched audio, its current accuracy is also its ceiling.
Whichever engine you choose, the differentiator is the evaluation and adaptation data behind it. Building that benchmark and adaptation set from natural bilingual speakers, to spec, is the part we supply.
Two sets. An evaluation set first: real calls or realistic re-enactments from your scenario, transcribed to a convention that marks the switch points, so you can score the agent where it fails. Then adaptation data, if the evaluation shows a gap the engine cannot close on its own: recordings by natural bilinguals in your register and vocabulary. How much depends on where the errors are (at the switch points, in the domain vocabulary, or in the acoustics), and a pilot batch is what tells you that. We scope the pilot to your call scenario before quoting a production line.
Tell us the languages, specs, and volume — we'll start with a pilot batch.
Start with a free pilot batch →