Collection of Taiwanese-language YouTube talk-show/documentary content, staged from COS ahead of a full transfer. Each COS source gets its own config (schemas differ, so they can't share one), each with one split named after the source folder.

Configs / Splits

ptv_ts_ds_test_zhtw (367 rows) -- PRELIMINARY, schema/field names copied as-is from ds/ptv_ts_ds_test_zhtw on COS, not yet normalized to this project's usual conventions (see rule.md). Label quality is unconfirmed; do not treat tw/zh as verified transcripts yet. A small test set sampled from 5 different YT shows (one episode each): 《大港ê台灣》, 《這是台灣款》, 《青春！咱的夢》, 無事坐巴士, 《台灣記事簿》. See origin for which show each row came from.
twasr_talking (9 rows) -- whole-episode audio + official diarization transcript, copied as-is from twasr_talking on COS (YT show "台灣新眼界"). One row per episode; each text is the complete *_dia_regular_txt.txt content, unparsed (speaker/timestamp lines not split out).
twasr_talking2 (151 rows) -- whole-episode audio with no transcript at all (twasr_talking2 on COS never had accompanying text), YT show "台灣新眼界．文化新台灣". id is the YouTube video ID.
ptv_ocr_test (1010 rows, 2.31 hours) -- utterance-level clips with cleaner OCR-derived Taiwanese/Mandarin text (no embedded Whisper timestamps, unlike ptv_ts_ds_test_zhtw), covering the same PTS YouTube shows. Sourced from a local staged copy (/home/wago/data/PTV_OCR_test), not directly from COS. Some clips' audio has Mandarin mixed in with the Taiwanese and this could not be cleanly separated out, so the whole split is still labeled/treated as Taiwanese audio overall -- treat text as "mostly Taiwanese, some Mandarin code-switching" rather than pure Taiwanese speech.

Dataset Structure

`ptv_ts_ds_test_zhtw`

idx: segment index within its source video (e.g. "0-8"), also used as the embedded audio.path filename. Not a globally unique ID across rows from different origin videos.
audio: audio clip.
tw / zh: Taiwanese / Mandarin text for the segment, with embedded Whisper-style timestamp tokens (e.g. <|0.00|>...<|2.06|>) not cleaned up.
origin: which source YT video this row came from (title + video ID).

`twasr_talking`

id: YouTube video ID.
audio: whole-episode audio clip (48kHz, stereo, 16-bit PCM WAV).
text: complete diarization transcript file content, unparsed.

`twasr_talking2`

id: YouTube video ID.
audio: whole-episode audio clip (16kHz, mono, 16-bit PCM WAV).

`ptv_ocr_test`

id: <video_id>_<segment_index>.
audio: audio clip.
text: Taiwanese OCR-derived transcript for the clip (the "language Taiwanese<asr_text>" prefix present in the raw source has been stripped). Audio may contain some code-switched Mandarin that couldn't be separated out -- see the note above.
mandarin: Mandarin OCR-derived transcript for the same clip (same prefix stripped).
origin: which source YT video this row came from (video ID).

See ds_raw_schema.md and corpus_inventory.md in the project repo for more context on where this came from and open questions about label quality.

Statistics

split	lang_name	hours	n_utts	n_chars_in_utts	secs/utt	chars/sec
ptv_ts_ds_test_zhtw	Taigi/Mandarin	2.1053	367	85221	20.65	11.24
twasr_talking	Taigi/Mandarin	8.3416	9	213055	3336.65	7.09
twasr_talking2	Taigi/Mandarin	131.522	151	0	3135.62	0
ptv_ocr_test	Taigi/Mandarin (code-switched)	2.3122	1010	34625	8.24	4.16
Total	-	144.2811	1537	332901	337.94	0.64

ptv_ts_ds_test_zhtw's n_chars_in_utts/chars_per_sec are computed on the raw tw column including the embedded <|timestamp|> tokens, so they're inflated compared to real speech rate. twasr_talking2 has no text at all (n_chars_in_utts/chars_per_sec are 0 because there's nothing to count, not because it's ASR-ready). ptv_ocr_test's n_chars_in_utts/chars_per_sec are computed on the text (Taiwanese) column.

Downloads last month: 50

Total file size:

17.7 GB