You need to agree to share your contact information to access this dataset

This repository is publicly accessible, but you have to accept the conditions to access its files and content.

Access requires agreeing to non-commercial research use only.

Log in or Sign Up to review the conditions and access this dataset content.

TAT-MOE

TAT (Taiwanese Across Taiwan) MOE 台文語音語料庫 -- speakers reading aloud example sentences from the Ministry of Education's official Taiwanese dictionary (臺灣台語常用詞 辭典), recorded simultaneously on 6 different microphones per sentence.

⚠️ Known incompleteness -- this is NOT the full official TAT-MOE corpus. The official release covers 328 train / 58 eval / 54 test speakers (86,072 / 16,357 / 15,962 sentences). What is actually available in the source COS bucket -- and therefore in this upload -- is smaller and unevenly distributed across institutions:

split official speakers speakers here official sentences sentences here
train 328 291 86,072 80,936
eval 58 47 16,357 13,269
test 54 45 15,962 13,087

Confirmed (2026-09-03) this is a genuine gap in the source data, not a mistake in how this repo was built: two institutions (宜蘭I / 宜蘭II) are entirely missing from train, and eval/test only contain the KH and TA institutions (every other institution -- IU, KN, SO, TH, TI, TS -- is completely absent from eval/test). Checked directly against the original zip archives' own file listings (not just what got extracted) and against an alternate 7z-packaged copy of the same data (moe/full/corpus-full.7z.*, byte-for-byte the same total size) -- neither contains the missing institutions either, so there is no more- complete source to recover this from within this project's COS bucket. If you need the full official corpus, re-source the missing institutions directly from the TAT-MOE release.

Dataset Structure

  • id: <speaker>@<sentence>_<rec_device> (e.g. TI_TIF1012@A067-5.2_condenser), unique per row (the same speaker/sentence appears once per rec_device, distinguished by this suffix).
  • audio: audio clip (16kHz mono PCM WAV).
  • text: Taiwanese (Han-lo) transcript, same language as the audio (originally tw in the source dataset).
  • mandarin: Mandarin transcript for the same sentence (originally zh).
  • rec_device: which of the 6 simultaneous microphones this row's audio came from.

Statistics

split lang_name hours n_utts n_chars_in_utts secs/utt chars/sec n_sents n_chars_in_sents
train Taigi 879.2111 485616 6625144 6.52 2.09 0 0
eval Taigi 145.761 79614 1259562 6.59 2.4 0 0
test Taigi 131.4756 78522 1006302 6.03 2.13 0 0
Total - 1156.4477 643752 8891008 6.47 2.14 0 0

train's stats combine an exact count for 272,775 rows (this run) with an estimate for the 212,841 rows uploaded before an earlier rate-limit crash (extrapolated from a 22-shard spread sample of those 256 pre-crash shards -- secs/utt/chars/sec were stable across the sample, so this should be accurate to within a percent or two). eval/test are exact (every row scanned in one continuous run).

Downloads last month
86