Taiwanese (Han-lo script) recitation readings from three combined sources, aligned via a VAD (silero) + faster-whisper + jiwer forced-alignment pipeline against official reference text (not OCR, not free-form YouTube captions):

jen-pin (jen-pin.com.tw): the site's own multi-year archive (民國 109-113) of Taiwan's national language recitation competition (全國語文競賽閩南語朗讀)'s officially authorized reading texts + audio, across all age/category groups.
knsh: a separately-recorded elementary-school recitation set, plus its own copies of the 112-year competition's officially labeled files.
llotg ("Lan-Lai-Oh-Taigi" / 咱來學臺灣閩南語): Taiwan's Ministry of Education's own official Southern-Min textbook reading-aloud audio.

Because all three sources pair audio against an officially authorized reference text (not ASR guesses or OCR), text quality is generally cleaner than this project's YouTube-scraped corpora -- spot-checked samples show clean Han-lo script with no obvious garbling.

Dataset Structure

id: {origin}_{idx}, built here (no id field in the source) -- verified unique across all rows.
audio: audio clip (16kHz, mono, 16-bit PCM WAV).
text: Taiwanese transcript, Whisper-style inline timestamp tokens stripped (same language/script as the audio, per this project's field-naming convention).
text_timestamped: the same transcript with the original inline Whisper timestamp tokens (<|0.00|>...<|9.98|>...) preserved untouched -- kept for potential future timestamp-supervised ASR fine-tuning (this is literally Whisper's own training target format, so no restructuring is needed to use it directly for that).
origin: which of the 196 source recordings/recitations this segment came from (e.g. "國中組_騎鐵馬", "0401-12" for an MOE textbook lesson, or a "...112全國語文競賽閩南語朗讀..."-style competition filename). Not globally unique on its own (a single origin has multiple segments); id is origin + idx and is unique.

Statistics

split	lang_name	hours	n_utts	n_chars_in_utts	secs/utt	chars/sec	n_sents	n_chars_in_sents
train	Taigi	14.001	2333	139270	21.6	2.76	0	0
Total	-	14.001	2333	139270	21.6	2.76	0	0

Downloads last month: 18

Use case

: ASR (Automatic Speech Recognition)

Language

: Taiwanese (Southern Min), Han-lo (漢羅) script

Source

arrow-format export on the source COS bucket) into HF-viewer-friendly Parquet shards.

Audio format

: 16kHz, mono, 16-bit PCM WAV