Datasetpublic

tiennguyenbnbk/gopt-vh-gold

VuiHoc GOPT Gold — audio + consensus labels (Arrow) 6,361 cau IELTS read-aloud thuan sach (~35h) tu he thong VuiHoc, moi row gom audio + nhan dong thuan 3 vendor (SpeechAce, SpeechSuper, iFlytek ISE), thang nghiep vu [0.0, 100.0], chia 4 split zero-leakage. Splits Split Mau Speakers Phone valid Word valid train 4,643 812 94.25% 98.25% val 581 100 94.28% 98.32% test_unseen_speakers 580 94 94.2% 98.13% test_unseen_prompts 557 190 93.2% 97.77%… See the full description on the dataset page: https://huggingface.co/datasets/tiennguyenbnbk/gopt-vh-gold.

sourceHugging Faceupdated 1mo agoView on Hugging Face

0likes43downloads

Dataset Card

VuiHoc GOPT Gold — audio + consensus labels (Arrow)

6,361 cau IELTS read-aloud thuan sach (~35h) tu he thong VuiHoc, moi row gom audio + nhan dong thuan 3 vendor (SpeechAce, SpeechSuper, iFlytek ISE), thang nghiep vu [0.0, 100.0], chia 4 split zero-leakage.

Splits

Split	Mau	Speakers	Phone valid	Word valid
train	4,643	812	94.25%	98.25%
val	581	100	94.28%	98.32%
testunseenspeakers	580	94	94.2%	98.13%
testunseenprompts	557	190	93.2%	97.77%

—Speaker khong giao nhau giua train / val / testunseenspeakers.
—testunseenprompts dung 7 de bai hoan toan moi (speaker co the trung train).

Schema (moi row)

Cot	Kieu	Mo ta
`id`	string	ID audio VuiHoc
`audio`	Audio	mp3 goc
`text`	string	de bai canonical
`user_id` / `question_id`	int64	nguoi hoc / de bai
`duration_sec`	float32	do dai audio
`split_role`	string	y nghia kiem thu cua split
`phone_list`	[39] string	bang tra ARPA39: index trong `phn`/`msdd_sub` -> ten am vi
`utt_heads`	[4] string	thu tu head utterance: accuracy, completeness, fluency, total
`word_heads`	[1] string	head word: accuracy
`scale`	string	thang diem: 0-100
`max_len`	int32	do dai pad cua cac sequence
`phn`	[150] int16	id ARPA39 canonical, pad = -1
`phone_label` / `phone_weight`	float32	diem am vi dong thuan [0,100] / trong so tin cay [0,1]; mask -1.0
`n_vendors`	uint8	so voter tai vi tri (0-3)
`word_id`	int16	chi so tu (trong text) de word-pooling, pad = -1
`word_acc` / `word_weight`	float32	word accuracy native vendor / trong so; mask -1.0
`msdd_type`	int16	0=OK, 1=Substitution, 2=Deletion, -1=Mask
`msdd_sub`	int16	id am vi thay the (dong thuan Ace∩SS), mask = -1
`utt_label` / `utt_weight`	[[4]] float32	(accuracy, completeness, fluency, total) / trong so
`utt_nv`	[[4]] uint8	so voter cua tung head

Quy uoc quan trong

—-1 / -1.0 la MASK (khong co nhan), KHONG phai diem so — loc bang (label >= 0).
—Phone: Median-of-3 vendor sau chuan hoa z-score ve khong gian SpeechSuper; 2 vendor chi nhan khi |dz| <= 1; nv <= 1 hoac bat dong -> mask.
—Fluency bi mask tren mot phan mau (2 vendor cham lech > 1 sigma).

GOP feature (se cap nhat)

Cot feat (CTC-GOP 80-d KoelLabs/xlsr-english-01 + occupancy) CHUA co trong version nay — se extract tu cot audio va day them len repo sau.

Load

python

from datasets import load_dataset
ds = load_dataset("tiennguyenbnbk/gopt-vh-gold")
ex = ds["train"][0]
ex["phone_list"], ex["utt_heads"]   # bang tra nam ngay trong row