hak high-quality speech corpus

This dataset contains the strict high-quality Hakka-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
hak	6,639,472	16,437.46

Audio subset overview

Counts below are unique packaged audio rows.

Audio subset	Unique rows	Approx. hours	Rejected rows	Parquet GiB
drama_wointian	50	0.09	0	0.01
hak_yijiazhizhu	18,636	145.83	0	9.63
hakka_cinema_001	153	0.26	0	0.01
hakka_cinema_002	125	0.19	0	0.01
hakka_cinema_003	166	0.18	0	0.01
hakka_cinema_005	116	0.14	0	0.01
hakka_cinema_007	355	0.47	0	0.03
hakka_cinema_008	89	0.20	0	0.01
hakka_cinema_010	39	0.07	0	0.01
hakka_cinema_011	258	0.34	0	0.02
hakka_cinema_017	18	0.02	0	0.00
hakka_cinema_019	22	0.04	0	0.00
hakka_cinema_022	17	0.02	0	0.00
hakka_cinema_023	33	0.07	0	0.00
hakka_cinema_027	45	0.09	0	0.01
hakka_cinema_029	88	0.15	0	0.01
hakka_cinema_031	56	0.07	0	0.00
hakka_cinema_040	146	0.26	0	0.01
hakka_cinema_045	66	0.15	0	0.01
hakka_cinema_049	129	0.28	0	0.02
hakka_cinema_052	120	0.42	0	0.03
hakka_processed_2023_year_hakka_collection	1,136	6.09	0	0.39
hakka_processed_2024_year_hakka_collection	4,316	24.51	0	1.55
hakka_processed_2025_year_hakka_collection	5,332	28.85	0	1.82
hakka_processed_2026_year_hakka_collection	1,854	9.54	0	0.60
hakka_processed_audio_fixed_25146	159,936	547.72	44	35.77
hakka_processed_eattogther	11,539	34.42	0	2.15
hakka_processed_show_hou	25,647	94.94	0	6.70
hakka_processed_singing	26,967	125.84	0	8.49
hakka_processed_whotoeat	10,187	59.38	0	3.91
hakka_tts/concat/dapu	299,221	1,242.95	0	53.74
hakka_tts/concat/hailu	293,221	1,231.23	0	53.88
hakka_tts/concat/nansixian	186,242	807.67	0	35.58
hakka_tts/concat/raoping	270,414	1,193.33	0	51.85
hakka_tts/concat/sixian	298,025	1,239.45	0	54.10
hakka_tts/concat/zhaoan	245,826	1,048.36	0	46.34
hakka_tts/elearning/dapu	349,296	501.05	0	27.45
hakka_tts/elearning/hailu	302,646	446.24	0	24.60
hakka_tts/elearning/raoping	295,273	460.70	0	25.47
hakka_tts/elearning/sixian	298,381	429.52	0	23.64
hakka_tts/elearning/zhaoan	286,811	424.81	0	23.52
hakka_tts/moe/dapu	527,551	1,003.67	0	56.80
hakka_tts/moe/hailu	564,338	1,083.34	0	62.22
hakka_tts/moe/nansixian	541,990	1,019.18	0	58.33
hakka_tts/moe/raoping	548,711	1,137.80	0	64.85
hakka_tts/moe/sixian	565,707	1,068.13	0	61.26
hakka_tts/moe/zhaoan	452,344	878.07	0	50.73
上家下屋	5,770	13.90	0	0.98
客家戲曲	7,460	21.96	0	1.24
客庄好味道	10,667	35.04	0	2.54
巷弄裡的吉光片羽	3,595	9.97	0	0.65
暗香風華	13,553	43.67	0	2.78
溜麵線雜貨店	2,744	9.77	0	0.63
維基百客	2,045	7.01	0	0.44

JSONL examples

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

{"audio_filepath":"audio/hakka_processed_2023_year_hakka_collection/27a6a418b124331d944e/part-00000.parquet#row=0","text":"觀眾朋友大家好，歡迎收看這禮拜的客家新聞雜誌，我是劉宜頻。延續之前我們的看見南非系列報導，今天我們再進一步關注被稱為彩虹之國的南非。雖然自然風景美不勝收，氣候宜人，有著不錯的先天生活環境，但事實上，不少移居南非的民眾日子卻是過得心驚膽戰。","context":"","duration":25.659,"lang":"<|HAK|>","audio_id":"77c489c16d0b955e06009e24caff44f01c233136","source_set":"hakka_processed_2023_year_hakka_collection","audio_format":"flac"}
{"audio_filepath":"audio/hakka_processed_2023_year_hakka_collection/959ae96148e15884dc9a/part-00000.parquet#row=0","text":"之所以有這樣的安全疑慮，是因為在南非產業結構明顯失衡，造成當地貧富差距愈來愈大；而在新冠疫情之後，貧窮、失業，種種國內經濟情勢蕭條加劇，釀成了南非幾乎每天都發生明目張膽的謀殺、搶劫等種種悲劇。客家新聞團隊記者胡吰誌跟林柏均前進南非的報導。","context":"觀眾朋友大家好，歡迎收看這禮拜的客家新聞雜誌，我是劉宜頻。延續之前我們的看見南非系列報導，今天我們再進一步關注被稱為彩虹之國的南非。雖然自然風景美不勝收，氣候宜人，有著不錯的先天生活環境，但事實上，不少移居南非的民眾日子卻是過得心驚膽戰。","duration":27.694,"lang":"<|HAK|>","audio_id":"1770997ed7f8c54e52f772262f36d32962936657","source_set":"hakka_processed_2023_year_hakka_collection","audio_format":"flac"}

Download only selected subsets

from huggingface_hub import snapshot_download

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

Or load an individual configuration:

from datasets import load_dataset
corpus = load_dataset("ACCOUNT/hak", "hak")
audio = load_dataset("ACCOUNT/hak", "audio-drama_wointian")

Release totals

• Unique audio: 6,639,472
• Audio hours: 16,437.460
• Parquet size: 0.917874 TB (0.834802 TiB)

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