OpenFormosa/common_voice_25_zh-TW

Common Voice Scripted Speech 25.0 - Chinese (Taiwan)

This repository mirrors the Chinese (Taiwan) (zh-TW) portion of Mozilla Common Voice Scripted Speech 25.0 in Hugging Face datasets format. The audio has been embedded into Parquet and exposed as a Hugging Face Audio feature at 48 kHz.

Official source page: Mozilla Data Collective - Common Voice Scripted Speech 25.0 - Chinese (Taiwan)

Dataset Details

Field	Value
Dataset ID	cmn2g7eaj01fio10769r1m96n
Common Voice release	cv-corpus-25.0-2026-03-09
Language	Chinese (Taiwan), 華語（台灣）, zh-TW
Spoken variety	Taiwan Mandarin / 中華民國國語, cmn-TW
Task	Automatic Speech Recognition (ASR)
Original format	MP3
Original archive	common-voice-scripted-speech-25-0-chines-e84858c5.tar.gz
Original archive size	2.95 GB
Release date on Mozilla Data Collective	2026-03-23
License	Creative Commons Zero v1.0 Universal (CC0-1.0)

The original Mozilla metadata describes this as a collection of read speech recordings in Chinese (Taiwan). It contains 140,630 clips from 2,317 self-selected volunteer speakers, totaling 131.33 hours of speech. Of those clips, 85,324 are validated, corresponding to 79.68 validated hours. The sentence pool contains 21,763 sentences.

Important Usage Notes

The source dataset is released under CC0-1.0. The Mozilla Data Collective page also lists additional restrictions and constraints:

• Do not attempt to determine the identity of speakers in Common Voice datasets.
• The source page states that re-hosting or re-sharing the dataset is forbidden.
• Use of this mirror should follow the Mozilla Data Collective dataset page and its current terms.

Speaker demographic fields are self-reported and optional. Blank values mean the speaker did not provide that field.

Loading

from datasets import load_dataset

ds = load_dataset("OpenFormosa/common_voice_25_zh-TW", split="train")
print(ds.features["audio"])
print(ds[0]["sentence"])

For quick inspection without downloading full splits:

from datasets import load_dataset, Audio

stream = load_dataset("OpenFormosa/common_voice_25_zh-TW", split="train", streaming=True)
stream = stream.cast_column("audio", Audio(sampling_rate=48000, decode=False))
row = next(iter(stream))
print(row["audio"])
print(row["sentence"])

Splits

The train, validation, and test splits are the official Common Voice training subsets. The validated, invalidated, and other splits preserve the broader Common Voice clip buckets. The training subsets are drawn from the validated bucket, so they should not be treated as disjoint from validated.

Split	Rows	Notes
train	7,394	Official training split
validation	5,119	Original dev.tsv
test	5,119	Official test split
validated	85,324	Validated clips
invalidated	4,920	Clips that did not pass validation
other	50,386	Clips pending validation or otherwise outside validated/invalidated

Mozilla's metadata reports 17,632 train/dev/test clips, covering 20.7% of validated clips. The average clip duration is 3.362 seconds.

Features

Column	Type	Description
audio	Audio(sampling_rate=48000)	Embedded MP3 audio exposed through the Hugging Face Audio feature
client_id	string	Hashed speaker identifier from Common Voice
path	string	Original relative MP3 filename
sentence_id	string	Common Voice sentence identifier
sentence	string	Expected transcription text
sentence_domain	string	Sentence domain label; may be blank or contain multiple comma-separated domains
up_votes	int32	Number of validators who accepted the clip
down_votes	int32	Number of validators who rejected the clip
age	string	Self-reported speaker age bucket
gender	string	Self-reported speaker gender
accents	string	Self-reported accent or place-of-origin label
variant	string	Language variant, if provided
locale	string	Locale code, normally zh-TW
segment	string	Custom dataset segment, if provided
duration_ms	int32	Clip duration from clip_durations.tsv

The original source files are also included under source/:

• source/README.zh-TW.md
• source/train.tsv
• source/dev.tsv
• source/test.tsv
• source/validated.tsv
• source/invalidated.tsv
• source/other.tsv
• source/reported.tsv
• source/validated_sentences.tsv
• source/unvalidated_sentences.tsv
• source/clip_durations.tsv

Demographics

Gender

Code	Label	Clips	Speakers
male_masculine	Male, masculine	68,527 (48.7%)	598 (25.8%)
female_feminine	Female, feminine	31,056 (22.1%)	258 (11.1%)
transgender	Transgender	100 (0.1%)	1 (0.0%)
non-binary	Non-binary	0	0
do_not_wish_to_say	Prefer not to say	25 (0.0%)	2 (0.1%)
Unspecified	Not declared	40,922 (29.1%)	1,575 (68.0%)

Gender declared: 99,708 of 140,630 clips (70.9%), 742 of 2,317 speakers (32.0%).

Age

Code	Label	Clips	Speakers
teens	Teens	8,440 (6.0%)	82 (3.5%)
twenties	Twenties	41,664 (29.6%)	451 (19.5%)
thirties	Thirties	27,021 (19.2%)	231 (10.0%)
fourties	Fourties	12,771 (9.1%)	105 (4.5%)
fifties	Fifties	12,587 (9.0%)	27 (1.2%)
sixties	Sixties	431 (0.3%)	3 (0.1%)
seventies	Seventies	30 (0.0%)	4 (0.2%)
eighties	Eighties	0	0
nineties	Nineties	0	0
Unspecified	Not declared	37,686 (26.8%)	1,540 (66.5%)

Age declared: 102,944 of 140,630 clips (73.2%), 777 of 2,317 speakers (33.5%).

Accent Distribution

The source page reports self-declared accent or place-of-origin coverage as follows.

Code	Label	Clips	Speakers
taipei_city	出生地：臺北市	19,646 (14.0%)	107 (4.6%)
new_taipei_city	出生地：新北市	8,850 (6.3%)	62 (2.7%)
taichung_city	出生地：臺中市	4,411 (3.1%)	47 (2.0%)
kaohsiung_city	出生地：高雄市	3,266 (2.3%)	42 (1.8%)
taoyuan_city	出生地：桃園市	3,015 (2.1%)	23 (1.0%)
hsinchu_city	出生地：新竹市	2,866 (2.0%)	11 (0.5%)
yunlin_county	出生地：雲林縣	2,560 (1.8%)	8 (0.3%)
nantou_county	出生地：南投縣	2,101 (1.5%)	7 (0.3%)
changhua_county	出生地：彰化縣	2,009 (1.4%)	22 (0.9%)
tainan_city	出生地：臺南市	1,708 (1.2%)	21 (0.9%)
chiayi_city	出生地：嘉義市	1,195 (0.8%)	5 (0.2%)
pingtung_county	出生地：屏東縣	913 (0.6%)	6 (0.3%)
hualien_county	出生地：花蓮縣	878 (0.6%)	5 (0.2%)
yilan_county	出生地：宜蘭縣	765 (0.5%)	8 (0.3%)
hong_kong	香港	690 (0.5%)	26 (1.1%)
chiayi_county	出生地：嘉義縣	379 (0.3%)	7 (0.3%)
hsinchu_county	出生地：新竹縣	343 (0.2%)	8 (0.3%)
keelung_city	出生地：基隆市	141 (0.1%)	10 (0.4%)
kinmen_county	出生地：金門縣	55 (0.0%)	1 (0.0%)
penghu_county	出生地：澎湖縣	20 (0.0%)	2 (0.1%)
miaoli_county	出生地：苗栗縣	15 (0.0%)	2 (0.1%)
taitung_county	出生地：臺東縣	10 (0.0%)	2 (0.1%)
Other	Other accent strings	5,017 (3.6%)	21 (0.9%)

Sentence Corpus

Most Traditional Chinese text was organized through the MozTW CC0 sentence corpus. Mozilla's metadata reports:

• Validated sentences: 20,786
• Unvalidated sentences: 977
• Pending sentences: 137
• Rejected sentences: 840
• Reported sentences: 179
• Total sentence pool: 21,763

Sentence source distribution:

Source	Sentences
sentence-collector	15,566 (74.9%)
setences	2,897 (13.9%)
MozTW CC0 corpus commit 01033097...	666 (3.2%)
MozTW CC0 corpus commit e340b6d...	451 (2.2%)
taipei_city_gov	355 (1.7%)
chatlogs	309 (1.5%)
Other	533 (2.6%)

Sentence Domains

The source page reports the following domain coverage among clips. In this Hugging Face version, sentence_domain is preserved as the raw Common Voice field and may be empty or contain multiple comma-separated values.

Code	Domain	Clips	Speakers
general	General	1,502 (1.1%)	84 (3.6%)
agriculture_food	Agriculture and Food	12 (0.0%)	7 (0.3%)
automotive_transport	Automotive and Transport	278 (0.2%)	45 (1.9%)
finance	Finance	3 (0.0%)	3 (0.1%)
service_retail	Service and Retail	151 (0.1%)	36 (1.6%)
healthcare	Healthcare	25 (0.0%)	18 (0.8%)
history_law_government	History, Law and Government	170 (0.1%)	39 (1.7%)
media_entertainment	Media and Entertainment	170 (0.1%)	44 (1.9%)
nature_environment	Nature and Environment	14 (0.0%)	12 (0.5%)
news_current_affairs	News and Current Affairs	44 (0.0%)	19 (0.8%)
technology_robotics	Technology and Robotics	777 (0.6%)	49 (2.1%)
language_fundamentals	Language Fundamentals	8 (0.0%)	7 (0.3%)

Intended Uses

Mozilla lists the intended use as training and evaluating automatic speech recognition models. The dataset may also be useful for computer-aided language learning and language or heritage revitalization applications.

Provenance and Acknowledgements

The text corpus was created by Mozilla Taiwan community contributors, the g0v community, and other open-source volunteers. Speakers are primarily individual volunteers from Taiwan.

Mozilla's metadata names Irvin Chen as the MozTW community contact for dataset table preparation.

License

The dataset is released under Creative Commons Zero v1.0 Universal. Users should also review the Mozilla Data Collective dataset page for current restrictions and terms before use.