README.md

metadata

license: cc-by-nc-sa-4.0
language:
  - zh
pretty_name: Minspeech (Southern Min)
extra_gated_prompt: >-
  Audio is sourced from copyrighted TV drama content and is released here
  strictly for non-commercial research use under CC BY-NC-SA 4.0, consistent
  with the original Minspeech license (https://minspeech.github.io/). By
  requesting access you agree to use this data only for non-commercial research
  purposes and not to redistribute the raw audio.
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
      - split: dev
        path: data/dev-*.parquet
      - split: test
        path: data/test-*.parquet
      - split: unlabeled
        path: data/unlabeled-*.parquet
dataset_info:
  features:
    - name: id
      dtype: string
    - name: audio
      dtype: audio
    - name: mandarin
      dtype: string

Minspeech (Southern Min)

Full train/dev/test transfer of the Minspeech Southern Min (Taiwanese) speech corpus, staged from an internal COS copy. unlabeled is still just a 10-clip illustrative preview -- see "Splits" below.

Dataset Description

• Use case: ASR (Automatic Speech Recognition)
• Language: spoken audio is Southern Min / Taiwanese (drama dialogue); the transcript (mandarin) is written in Mandarin Chinese, not Taiwanese Han-lo -- the two are different languages, matching the official Minspeech release as-is.
• Source: Minspeech official corpus (Southern Min TV drama), re-cropped to match the official splits exactly. See "Splits" below for details.
• Audio format: 16kHz, mono, PCM WAV
• Split: 4 splits, see below.

Dataset Structure

• id: official Minspeech utterance ID (S<series><episode><utterance> for train/dev/ test; source video ID for unlabeled).
• audio: audio clip (16kHz mono WAV), spoken in Southern Min / Taiwanese.
• mandarin: Mandarin Chinese transcript, exactly as released in the official *_text files. Not a Taiwanese-script transcription of the audio. Empty for unlabeled (see below).

Splits

• train / dev / test: full official Minspeech ASR splits, utterance-level clips with transcripts, matching the official *_wav.scp / *_text files. One source episode (S01142) failed to parse and was skipped -- everything else is included.
• unlabeled: raw, not utterance-segmented source video audio with no transcript (mandarin is an empty string). The original Minspeech release ships only a video ID list for this portion and expects users to run their own VAD (the authors suggest silero-vad) to cut it into utterances. This is still just a 10-clip preview (20-second excerpts from the start of 10 source videos, for illustration only) -- the full unlabeled split (807GB of raw video audio) hasn't been uploaded yet.

Statistics

split	lang_name	hours	n_utts	n_chars_in_utts	secs/utt	chars/sec	n_sents	n_chars_in_sents
train	Taigi	1692.21	2614243	20301907	2.33	3.33	0	0
dev	Taigi	16.7259	25671	201198	2.35	3.34	0	0
test	Taigi	18.509	28685	219051	2.32	3.29	0	0
unlabeled	Taigi	0.0556	10	0	20.02	0	0	0
Total	-	1727.4995	2668609	20722156	2.33	3.33	0	0

dev/test/unlabeled are computed exactly (every row scanned). train's hours/ n_chars_in_utts are estimated from a spread sample of 48 of its 1186 shards (~4%) and scaled up to the exact row count (2,614,243, from the Hub's own row-count metadata) -- scanning all 1186 shards' audio in full wasn't done here since it would mean re-downloading ~178GB just to compute stats. secs/utt and chars/sec are stable across the sampled shards (2.3-2.4s/utt, ~3.3 chars/sec throughout), so the estimate should be accurate to within a percent or two, but treat the train row as an estimate, not an exact count.