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

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.

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

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

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

Age

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.

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:

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.

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.