Dataset Information
- Corpus:
hakka_elearning - Status:
Public (Gated) - Split Policy:
trainonly (in_ood_testcolumn preserved for traceability without injecting noisy pseudo-labels into eval/test metrics)
Subset Stats
| Subset | Train (Samples / Hours) | CER Before -> After | Eval (Samples / Hours) | Description |
|---|---|---|---|---|
clean_asr |
7,279 / 5.29 | 25.79% -> 9.71% | 0 / 0.00 | Recommended ASR training pool (clean + clean_colloquial, zero overlapping intervals) |
clean |
5,971 / 4.22 | 24.14% -> 6.21% | 0 / 0.00 | Strict gold subset (clean_exact + clean_revised, CER & IPA-CTC verified) |
clean_exact |
5,700 / 4.02 | 23.16% -> 6.18% | 0 / 0.00 | Verbatim original reference matches (text_edited=false) |
clean_revised |
271 / 0.20 | 39.50% -> 6.79% | 0 / 0.00 | Kaldi Biased-LM + Smith-Waterman revised/trimmed utterances (text_edited=true) |
clean_colloquial |
1,308 / 1.07 | 31.88% -> 22.46% | 0 / 0.00 | High CTC+LID colloquial utterances promoted from review_text |
review_text |
3,521 / 2.72 | 37.98% -> 29.09% | 0 / 0.00 | Audio QC passed, text flagged for human review (includes clean_colloquial) |
rejected |
6,534 / 4.28 | 80.84% -> 79.84% | 0 / 0.00 | Filtered out (noise, bleed conflict, Mandarin mismatch, alignment failure) |
all |
16,026 / 11.22 | 47.93% -> 38.55% | 0 / 0.00 | Complete processed corpus across all buckets (100% of utterances) |
Quick Load
from datasets import load_dataset
ds = load_dataset("formospeech/hakka_elearning_cleaned", "clean_asr", split="train")
ds_revised = load_dataset("formospeech/hakka_elearning_cleaned", "clean_revised", split="train")
- Downloads last month
- 1,462