FCU_Voice_47_EX

Preprocessed Mandarin tone recognition data for ToneMamba inference. Raw recordings are not required.

Dataset input

csv/reset/
├── all.csv
├── train_filtered.csv
├── val_filtered.csv
├── test_filtered.csv
├── train_dropped.csv
├── val_dropped.csv
└── test_dropped.csv
f0/
├── f0_praat.csv
└── f0_praat.npy

Split	Filtered CSV rows	Usable IPUs	Evaluated syllables
train	1,136	1,132	5,328
val	124	123	568
test	315	314	1,327

Usable counts apply the ToneMamba F0 and reference-label filters. An IPU is an inter-pausal unit. Tone labels are 1–5; 5 denotes neutral tone.

Fields

The split CSVs contain speaker and sample identifiers, syllable and word timing, pinyin_word, tone_word, duration_final, and duration_v. Match each split row's <filename>_<index_sentence> to filename in f0/f0_praat.csv; CSV row order corresponds to the contours in f0/f0_praat.npy. The F0 artifacts contain 4,556 contours.

Use the F0 metadata's speaker-level min and max for normalization and val_num for the finite-frame count. The split CSV's f0_min and f0_max columns are placeholders. Contours use a 10 ms frame step and NaN for unvoiced frames. The NPY contains a variable-length object array and is loaded with numpy.load(..., allow_pickle=True); load only trusted artifacts.

Download

from huggingface_hub import snapshot_download

path = snapshot_download(
    repo_id='poyu39/FCU_Voice_47_EX',
    repo_type='dataset',
    allow_patterns=['csv/reset/*', 'f0/*'],
)

Private-repository access requires an authorized Hugging Face account. ToneMamba automatically downloads the three filtered split CSVs and two F0 files, then reuses the local Hugging Face cache.