Datasets:
Notices
This dataset contains no audio. It contains acoustic measurements — numeric facts about recordings — and statistical aggregates over them. This file records where every measurement came from and on what terms, and is the authoritative statement; LICENSE is a summary.
The dataset itself — CC BY 4.0
The selection, measurement pipeline, feature definitions, templates, variability model, calibration and documentation are © 2026 Den Patin, licensed CC BY 4.0. Use it for anything, including commercially; attribute as:
twsyllables — Taiwanese Mandarin syllable acoustics, Den Patin,
https://huggingface.co/datasets/taiwan-corpora/twsyllables, CC BY 4.0
Anything redistributing this dataset or a derivative of it must also carry the 全字庫 attribution below.
Source recordings and their terms
| source | speakers in tokens |
license of the audio | what was taken |
|---|---|---|---|
| 教育部《國語辭典簡編本》 audio | moe_tw |
CC BY-ND 3.0 TW | measurements only |
| 當代中文課程 (A Course in Contemporary Chinese) narration | dangdai_tw |
proprietary | measurements only |
| 數位發展部 全字庫 (CNS 11643) syllable audio | cns_f, cns_m |
OGDL v1.0 | measurements only |
| Mozilla Common Voice 26.0 zh-TW | none (aggregates only) | CC0 1.0 | between-speaker variance widths |
全字庫 — OGDL v1.0
The 全字庫 audio is offered by the Ministry of Digital Affairs under 政府資料開放授權條款 第1版 (OGDL v1.0), which grants unrestricted use, adaptation and redistribution, including commercial, subject to attribution. Failure to attribute voids the grant ab initio. Attribution string:
數位發展部，CNS11643中文標準交換碼全字庫網站
MOE dictionary audio — CC BY-ND 3.0 TW
The recordings themselves are under a NoDerivatives license and are not included, adapted or redistributed here in any form. What this dataset publishes about them is measurements: durations, fundamental-frequency values, formant frequencies, spectral moments, energy levels. A measured quantity is a fact about a recording, not a reproduction or adaptation of it; Taiwan's Copyright Act (著作權法) art. 10-1 confines protection to expression, expressly excluding ideas, procedures, processes, systems, methods of operation, concepts, principles and discoveries — and facts are not expression. The 16-point contour summaries and resampled coefficient trajectories published here cannot reconstruct the audio, the voice, or any protectable expression. The same theory underlies published n-gram counts over copyrighted text. Attribution, given here as a courtesy and as scholarly practice:
教育部《國語辭典簡編本》，CC BY-ND 3.0 TW
當代中文課程 narration — proprietary
Same position as above: the recordings are not redistributed and cannot be
reconstructed; only measurements are published. The transcript field
reproduces the single words and characters being read — individual words are
not copyrightable subject matter.
Common Voice — CC0
Common Voice releases are CC0; no restriction applies. Only aggregate per-feature dispersion estimated across 271 speakers appears here — no per-speaker rows, no clip identifiers, no demographic fields. Speakers are identified upstream only by pseudonymous client IDs, and per Mozilla's terms no attempt to identify speakers has been or may be made.
Privacy
No personal data is included. The four token-corpus voices are institutional or professional recordings published by their institutions; Common Voice contributions are pseudonymous and appear only as aggregates.
Trademarks
教育部, 數位發展部, Mozilla and Common Voice are names of their respective institutions and projects; their appearance here is attribution, not endorsement.