Springleaf/Intoner-Dataset

Intoner Dataset

English · 简体中文

Data for Intoner: For Chinese Poetry Intoning Synthesis (IJCAI 2025). Code: Intoner.

• cfe/ — CFESet (paper §5.1): 10,765 annotated records with pinyin, tone, rhyme and caesura labels per character (manifest.csv; train.csv 9,689 / test.csv 1,076), plus the pronunciation dictionary dict_modify.txt. The separate eval_paper.csv contains the 213 poems used for paper Table 3.
• pis/ — PISSet (paper §5.2): 20,112 sentence-level intoning recordings by 24 intoners (22.05 kHz).
  • raw/<speaker>/<item>.wav audio and .lab transcripts
  • processed/ phoneme metadata, phoneme inventory, MFA dictionary and MFA TextGrid alignments
  • binary/ ready-to-train features (train / validation / test: 18,718 / 871 / 523)

Download into the code repository with:

uvx --from huggingface_hub hf download Springleaf/Intoner-Dataset --repo-type dataset --local-dir data/intoner

wav_fn in pis/processed/metadata_phone.csv and the binary records is relative to the dataset root, for example pis/raw/S002/S00200001_01.wav.

Evaluate the paper-specific CFE set from the code repository with:

uv run python scripts/eval_cfe.py --checkpoint checkpoints/cfe/model_1500.pt --csv data/intoner/cfe/eval_paper.csv --output outputs/cfe_eval_paper.json

cfe/polyphones.csv is the character list used for Table 3 polyphone accuracy, with columns 字|多音|出现音.

The output includes pinyin, polyphone (poly), tone, rhyme and stop accuracies; add --no-dictionary for CFE w/o D.

File formats are documented in docs/dataset.md of the code repository.

Citation

@inproceedings{zuo2025intoner,
  title     = {Intoner: For Chinese Poetry Intoning Synthesis},
  author    = {Zuo, Heda and Sun, Liyao and Lai, Zeyu and You, Weitao and Chen, Pei and Sun, Lingyun},
  booktitle = {Proceedings of the Thirty-Fourth International Joint Conference on Artificial Intelligence, {IJCAI-25}},
  pages     = {10252--10260},
  year      = {2025},
  doi       = {10.24963/ijcai.2025/1139}
}