README.md

metadata

license: apache-2.0
language:
  - zh
pretty_name: THCHS-30 Segment
task_categories:
  - automatic-speech-recognition
tags:
  - speech
  - phone-alignment
  - segmentation
  - mandarin
size_categories:
  - 10K<n<100K

THCHS-30 Segment

Mandarin Chinese read-speech corpus with phone-level time alignments. Suitable for training and evaluating phone recognition and phonetic segmentation models.

Sources

• Audio: THCHS-30 (OpenSLR 18) by Dong Wang, Xuewei Zhang, Zhiyong Zhang (Tsinghua University, 2015).
• Phone alignments: anyspeech/THCHS-30-alignments.

Splits

Split	Utterances
train	10,000
val	893
test	2,495

Splits follow the original OpenSLR 18 directory partition (data_thchs30/{train,dev,test}); dev is renamed to val.

Schema

Column	Type	Description
utt_id	string	Utterance id, e.g. A11_0
audio	Audio(16 kHz)	Embedded waveform bytes (decoded on access)
text	string	Hanzi sentence transcript
phones	sequence[string]	IPA phone tokens with tone diacritics
phone_starts	sequence[float64]	Phone start times in seconds
phone_ends	sequence[float64]	Phone end times in seconds
language	string	cmn (ISO 639-3)
speaker_id	string	Speaker code (utt_id prefix, e.g. A11)
duration	float64	Utterance duration in seconds
split	string	train / val / test

Phone inventory

Phones are IPA with Mandarin tone diacritics, e.g. lː, y˥˩, ʂ˘, ɻ̩˥˩, a˧˥˘. Silence and pauses are marked with [SIL] intervals, which are kept in the alignment so boundary models can learn from them.

License

Released under the Apache 2.0 license, matching the original THCHS-30 release.

Citation

@misc{THCHS30_2015,
  title={THCHS-30 : A Free Chinese Speech Corpus},
  author={Dong Wang, Xuewei Zhang, Zhiyong Zhang},
  year={2015},
  url={http://arxiv.org/abs/1512.01882}
}