Datasheet

This datasheet documents CQC Release 1.0, following the questions proposed in Datasheets for Datasets (Gebru et al., 2021).

Sections: Motivation · Composition · Collection · Preprocessing · Uses · Distribution · Maintenance

Motivation

CQC was created to support research on Mandarin speech, sex, gender, and sexuality. Many existing Mandarin speech corpora were not designed for these questions and record gender, if at all, as a binary label; research on gender, sexuality, and the voice has focused on English and a few other languages. CQC was built by the CQC Research Team for research purposes.

Composition

• 54 speakers, ≈10.5 hours of speech, 5,183 transcribed segments.
• Each speaker described themselves as a member of a sexual or gender minority community in their own public content.
• Speaker-level fields: gender identity, trans status, sex characteristics, sexual orientation, and Mandarin variety where stated. Fields not stated by the speaker are coded not stated.
• Session-level fields: speech genre, recording conditions.
• The corpus contains information about sexual orientation and gender identity, which is sensitive.

Collection

• Source: public videos on Bilibili and YouTube, found with Chinese and English search terms.
• Inclusion required explicit self-description, Mandarin speech, adequate recording quality, and at least five minutes of usable speech.
• Speakers were not contacted, and individual consent was not obtained. This is why the corpus is under controlled access, and why removal is available on request.
• The project was reviewed by the PI’s institutional review board.

Preprocessing

• Audio converted to 16 kHz mono WAV; two-speaker recordings diarized with pyannote.audio, with target-speaker segments checked by hand.
• Transcribed automatically with Whisper; all segments reviewed by trained annotators; 3,329 of 5,183 segments (64.2%, 95% CI 62.9–65.5%) accepted without change.
• Text converted to simplified characters, punctuation removed.
• Words and phones aligned automatically with the Montreal Forced Aligner (pretrained Mandarin model), then checked by hand.

Uses

Intended

Phonetic, sociophonetic, sociolinguistic, and social-scientific research; aggregate evaluation of speech technology across speakers.

Prohibited

Inferring or classifying identity from speech, text, or metadata; identifying or contacting speakers; voice cloning or synthesis; commercial use; redistribution. See the Data Use Agreement.

Distribution

Documentation and aggregate statistics are public on this website. Transcripts, TextGrids, metadata, timestamps, and audio files are provided only to approved researchers under the Data Use Agreement.

Maintenance

• Maintained by the CQC Research Team; contact chinesequeercorpus@outlook.com.
• Sources are re-checked before each release; material whose source was deleted or made private is removed.
• Removal requests are handled as described on the Ethics page; approved users are notified and must delete removed material.
• Every release is numbered and listed on the Corpus page with its changes.
• Known limitations are listed on the Corpus page.