nan high-quality speech corpus

This dataset contains the high-quality Minnan-to-Chinese S2TT corpus and its source-partitioned audio.

All audio is FLAC PCM16. Source sample rate and channel count are preserved. Every packaged item passed a source-to-FLAC bit-identical PCM round trip and a separate full-release decode/hash validation.

Corpus rows use portable locators of the form audio/<subset>/<group>/part-N.parquet#row=N.

Corpus overview

Corpus/config	Rows	Approx. hours
nan	2,943,134	9,172.24

Audio subset overview

Counts below are unique packaged audio rows.

Audio subset	Unique rows	Approx. hours	Rejected rows	Parquet GiB
HiHi導覽先生	14,033	61.83	0	4.05
bonchacha	13,702	38.38	0	2.51
minnan_crawl	333,148	103.92	37	6.35
minnan_crawl_2	106,568	127.16	0	9.18
minnan_gold_ambertalk	2,412	7.01	0	0.48
minnan_gold_be_ki_c86fddff	38	0.14	0	0.01
minnan_gold_billionairestory	136,732	380.04	0	27.53
minnan_gold_blue	13,966	39.81	0	2.55
minnan_gold_daaidrama	186,534	1,356.66	0	83.45
minnan_gold_daidai_friends_fbc8a6f8	17	0.05	0	0.00
minnan_gold_dowry	181,449	525.94	0	33.00
minnan_gold_dragondance	112,429	343.18	0	22.74
minnan_gold_fengshui	221,083	651.97	0	41.77
minnan_gold_formosa_tv_dramas_d891ae92	112,734	307.77	0	20.38
minnan_gold_gloryoflove	142,544	405.27	0	29.20
minnan_gold_goldencity	227,188	644.51	0	44.64
minnan_gold_greattimes	152,270	451.84	0	31.23
minnan_gold_happiness	134,236	390.89	0	26.26
minnan_gold_hope	1,163	3.18	0	0.24
minnan_gold_li_giok_hong_ddcfcdd9	76	0.27	0	0.02
minnan_gold_lucky	124,311	353.97	0	25.33
minnan_gold_ptstalk_b6580a36	242	0.84	0	0.05
minnan_gold_scarecrowcontemporary_dance_company_d4a0c873	1,737	4.15	0	0.24
minnan_gold_score	43,749	121.55	0	6.85
minnan_gold_set_04910977	529	1.52	0	0.08
minnan_gold_set_0cb7db62	226	0.73	0	0.04
minnan_gold_set_246f6897	450	1.35	0	0.08
minnan_gold_set_26788ebf	3,016	9.97	0	0.57
minnan_gold_set_34c0c58d	87	0.28	0	0.02
minnan_gold_set_46a5a4a3	4,649	13.25	0	0.92
minnan_gold_set_4a40fe6c	4,636	15.84	0	0.68
minnan_gold_set_62d47487	175	0.40	0	0.02
minnan_gold_set_792fa89b	479	1.23	0	0.07
minnan_gold_set_7e1b10d0	395	1.10	0	0.07
minnan_gold_set_889bc2dd	43	0.09	0	0.01
minnan_gold_set_98c098ad	73	0.21	0	0.01
minnan_gold_set_c24accde	4,958	17.11	0	0.87
minnan_gold_set_c9f74ff0	6,759	23.68	0	1.05
minnan_gold_set_cb85b17b	1,277	4.89	0	0.27
minnan_gold_set_drama_f4d63b33	36,337	99.50	0	7.27
minnan_gold_set_f138f50c	59	0.19	0	0.01
minnan_gold_springflower	180,687	532.92	0	34.72
minnan_gold_taiwanan	90,751	258.22	0	17.04
minnan_gold_taiwanannew	21,142	59.55	0	3.95
minnan_gold_taiwanlegend	29,537	85.50	0	5.84
minnan_gold_tat_collected	3,871	11.09	0	0.66
minnan_pseudo	253,102	1,572.70	0	104.48
tat_hua	4,447	16.20	0	1.11
this_is_taiwan	10,479	52.20	0	3.40
廣告start	1,745	6.87	0	0.45
萬秀豬王	9,190	26.89	0	2.14
豬哥會社	3,128	10.62	0	0.82
霹靂兵烽決	1,373	2.63	0	0.20
霹靂天機	994	2.34	0	0.19
黑白威廉	4,042	14.32	0	0.90
黑貓夜總會	2,137	8.28	0	0.55

JSONL examples

Two abridged corpus rows are shown below; the locators point directly to the packaged Parquet audio rows.

{"audio_filepath":"audio/minnan_gold_ambertalk/a148aad3e6406b73ece6/part-00000.parquet#row=2321","text":"今天帶大家來關渡宮。","context":"","duration":1.621,"lang":"<|NAN|>","audio_id":"bacc2a1efaa39393985d7dfc014cc288538a650b","source_set":"minnan_gold_ambertalk","audio_format":"flac"}
{"audio_filepath":"audio/minnan_gold_ambertalk/a148aad3e6406b73ece6/part-00000.parquet#row=1621","text":"不得了啦，不得了啦！","context":"今天帶大家來關渡宮。","duration":2.427,"lang":"<|NAN|>","audio_id":"ff309f9a1d4e045e7694dde0f26a60441bd7d9c1","source_set":"minnan_gold_ambertalk","audio_format":"flac"}

Download only selected subsets

from huggingface_hub import snapshot_download

snapshot_download(
    repo_id="ACCOUNT/nan",
    repo_type="dataset",
    allow_patterns=[
        "corpus/*",
        "audio/HiHi--d74df3ad55bc/*/part-*.parquet",
    ],
)

Or load an individual configuration:

from datasets import load_dataset
corpus = load_dataset("ACCOUNT/nan", "nan")
audio = load_dataset("ACCOUNT/nan", "audio-HiHi-d74df3ad55bc")

Release totals

• Unique audio: 2,943,134
• Audio hours: 9,172.002
• Parquet size: 0.651289 TB (0.592344 TiB)

See release.json for exact per-subset counts, hours, bytes, and corpus checksums.