Datasets:

formospeech
/

hat_asr_nansixian_reading_clean_r

dataset_info:
  config_name: Hakka_NanSixian
  features:
    - name: id
      dtype: string
    - name: duration
      dtype: float64
    - name: hanzi
      dtype: string
    - name: hanzi_cln
      dtype: string
    - name: pinyin
      dtype: string
    - name: pinyin_cln
      dtype: string
    - name: ipa
      dtype: string
    - name: ipa_cln
      dtype: string
    - name: mandarin
      dtype: string
    - name: lang_group
      dtype: string
    - name: lang_group_en
      dtype: string
    - name: speaker
      dtype: string
    - name: mismatched_trs
      dtype: bool
    - name: audio
      dtype: audio
  splits:
    - name: train
      num_examples: 36070
configs:
  - config_name: Hakka_NanSixian
    data_files:
      - split: train
        path: Hakka_NanSixian/train-*
task_categories:
  - text-to-speech
language:
  - zh
  - hak
size_categories:
  - 10K<n<1M

hat_asr_nansixian_reading_clean_r

This dataset is an enhanced -R variant of formospeech/hat_asr_nansixian_reading_clean.

Summary

Subset: Hakka_NanSixian
Dialect: 客語南四縣
Train samples: 36070
Audio: enhanced 24 kHz WAV

TRAIN

Subset	lang_group	hours	n_utts	n_chars	secs/utt	chars/sec
Hakka_NanSixian	客語_南四縣	71.49	36,070	2,046,782	7.14	7.95
Total	-	71.49	36,070	2,046,782	7.14	7.95

Processing

Start from the original formospeech/hat_asr_nansixian_reading_clean train split (16 kHz native).
Run speech enhancement with nvidia/RE-USE (Multilingual Universal Speech Enhancement), using its bandwidth-extension mode (--BWE 24000) to enhance and upsample to 24 kHz in one pass.
Replace only the audio field with the enhanced audio; all other fields are copied unchanged from the source, one row per source row (no rows dropped or added).
Encode as 16-bit PCM WAV.

Why RE-USE (changed from the previous `sidon-v0.1`-based version)

The previous version of this dataset (see below) used sarulab-speech/sidon-v0.1 for enhancement. That model was observed to occasionally distort pronunciation or alter speaker timbre on this corpus. nvidia/RE-USE's own technical report (arXiv:2603.02641, §3.10) directly validates downstream TTS training on RE-USE-enhanced audio, reporting improved CER/WER and speaker-similarity (not degraded) versus unenhanced audio -- the closest available evidence to this dataset's actual use case, rather than generic audio-quality benchmarks alone.

Previous version (sidon-v0.1-based)

The prior revision of this dataset -- produced with sarulab-speech/sidon-v0.1 instead of RE-USE -- remains accessible via its commit hash for reproducibility (e.g. if a model was trained on that version and needs to be reproduced exactly):

from datasets import load_dataset

ds = load_dataset(
    "formospeech/hat_asr_nansixian_reading_clean_r",
    revision="962a18e183f4b0e7ea7d8167ac9d7dd4281c270a",
)

or with huggingface_hub:

from huggingface_hub import snapshot_download

snapshot_download(
    repo_id="formospeech/hat_asr_nansixian_reading_clean_r",
    repo_type="dataset",
    revision="962a18e183f4b0e7ea7d8167ac9d7dd4281c270a",
)