Dataset Information
- Corpus:
hakkaradio_news - 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 |
47,566 / 162.12 | 18.33% -> 18.03% | 0 / 0.00 | Recommended ASR training pool (clean + clean_colloquial, zero overlapping intervals) |
clean |
25,674 / 83.10 | 13.22% -> 12.54% | 0 / 0.00 | Strict gold subset (clean_exact + clean_revised, CER & IPA-CTC verified) |
clean_exact |
24,517 / 81.24 | 12.75% -> 12.60% | 0 / 0.00 | Verbatim original reference matches (text_edited=false) |
clean_revised |
1,157 / 1.86 | 29.38% -> 9.89% | 0 / 0.00 | Kaldi Biased-LM + Smith-Waterman revised/trimmed utterances (text_edited=true) |
clean_colloquial |
24,137 / 86.50 | 22.84% -> 22.84% | 0 / 0.00 | High CTC+LID colloquial utterances promoted from review_text |
review_text |
55,056 / 185.76 | 28.11% -> 28.11% | 0 / 0.00 | Audio QC passed, text flagged for human review (includes clean_colloquial) |
rejected |
9,553 / 21.68 | 52.96% -> 52.96% | 0 / 0.00 | Filtered out (noise, bleed conflict, Mandarin mismatch, alignment failure) |
all |
90,283 / 290.53 | 25.49% -> 25.32% | 0 / 0.00 | Complete processed corpus across all buckets (100% of utterances) |
Quick Load
from datasets import load_dataset
ds = load_dataset("formospeech/hakkaradio_news_cleaned", "clean_asr", split="train")
ds_revised = load_dataset("formospeech/hakkaradio_news_cleaned", "clean_revised", split="train")
- Downloads last month
- 5,898