README.md

metadata

license: unknown
language:
  - zh
pretty_name: iFormosa
extra_gated_prompt: Access requires agreeing to non-commercial research use only.
extra_gated_fields:
  Name: text
  Affiliation: text
  Intended use: text
  I agree to use this dataset only for non-commercial research purposes: checkbox
configs:
  - config_name: default
    data_files:
      - split: train
        path: data/train-*.parquet
dataset_info:
  features:
    - name: id
      dtype: string
    - name: audio
      dtype: audio
    - name: text
      dtype: string
    - name: text_timestamped
      dtype: string
    - name: mandarin
      dtype: string
    - name: mandarin_timestamped
      dtype: string

iFormosa

iFormosa（閩南語文數位教材）Southern Min digital teaching material -- read-aloud passages with Taiwanese (Han-lo) transcripts and Mandarin translations.

Dataset Description

• Use case: ASR (Automatic Speech Recognition)
• Language: Taiwanese (Southern Min), Han-lo (漢羅) script
• Source: reshaped from ds/iformosa-ts-ds.zip (a datasets arrow-format export on the source COS bucket) into HF-viewer-friendly Parquet shards. The original source (scraps/oneai/iformosa/) shows richer per-word annotation (Han-lo + Tai-lo romanization per word) than what made it into this arrow export; only the whole-sentence tw/zh pair survived into this repo.
• Known content overlap: at least one passage ("阿公，行啦！咱來去騎鐵馬...") is identical to an entry in NickWeng/llotg-jenpin-knsh (its jen-pin/"國中組_騎鐵馬" source) -- the two corpora likely draw from the same underlying reading-text bank in places. Not deduplicated across repos; if you're combining both for training, check for duplicate text/audio content.
• Audio format: 16kHz, mono, 16-bit PCM WAV
• Split: this dataset is train only.

Dataset Structure

• id: source utterance identifier (already unique, e.g. 1090101_0-3).
• 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 preserved untouched.
• mandarin: Mandarin translation of the same passage. Empty for 97 of 316 rows (no translation was available in the source for those passages -- not a data error).
• mandarin_timestamped: the same, with original inline timestamp tokens preserved (also empty where mandarin is empty).

Statistics

split	lang_name	hours	n_utts	n_chars_in_utts	secs/utt	chars/sec	n_sents	n_chars_in_sents
train	Taigi	2.0955	316	19575	23.87	2.59	0	0
Total	-	2.0955	316	19575	23.87	2.59	0	0