Speech data for Taiwanese Hokkien and Hakka, with consent on every file.
Natural two-channel conversations and private test sets for speech-model teams. Recorded in Taiwan, including the Mandarin–Hokkien–English switching of real speech.
What we provide
Private test sets
Held-back conversational audio your models have never seen. We run your model and send an accuracy report by variety, age group and code-switching.
Conversational datasets
Spontaneous two-person conversations, one channel per speaker, with transcripts and full labels. Non-exclusive licenses.
Custom collection
Specific varieties, settings or hours on request: Hakka Sixian, Hailu, Dapu, Raoping, Zhao'an, and Hokkien beyond Taiwan.
Why it's different
Spontaneous, not read
Public Taiwanese corpora total about 300 hours and are mostly read speech. Models fail most on natural conversation.
Consent you can audit
Every speaker signs consent in the app. Each file carries a record of who, when and under which license, for EU AI Act training-data disclosures.
Privacy-scrubbed
Names, phone numbers, addresses and ID numbers spoken aloud are removed from audio and tagged in transcripts.
Data spec
| Audio | 48 kHz, 16-bit WAV (original rate kept), one channel per speaker; 16 kHz copies on request |
|---|---|
| Transcripts | MOE Han characters plus Tâi-lô (Hokkien) or Taiwan Hakka Pinyin, with Mandarin translation; segment timestamps |
| Labels | Variety and sub-variety, hometown county, age band, first or second language, read or spontaneous, code-switch tags per segment, setting |
| Quality | Automatic noise, clipping, language and duplicate-speaker checks; synthetic-voice screening; native reviewer on every transcript; 5% double review |
| Delivery | Secure download links; JSON manifest with a consent and origin record per file |
| License | Non-exclusive, commercial AI training and evaluation; summary |
Request a sample or an evaluation
Tell us what you need. You'll get an automatic reply now and a personal answer within one business day.