Corpus

CQC Release 1.0 contains about 10.5 hours of Mandarin speech from 54 speakers, drawn from public Bilibili and YouTube videos, transcribed and time-aligned at the level of words and individual speech sounds.

On this page: Key terms · Overview · Sources and inclusion · Metadata and annotation · Processing · Data quality · Statistics · Limitations · Releases

Key terms

This page uses a few technical terms. Each is explained once here.

Speech corpus	A collection of recordings, organized and annotated for research.
Transcription	Writing down what is said in a recording, word for word.
Phone	A single speech sound, such as the m or the a in 妈 mā.
Forced alignment	Automatically marking where each word and each speech sound begins and ends in the recording.
TextGrid	The file format that stores these time marks; it opens in the free program Praat.
Speaker diarization	Automatically working out who is speaking when, in a recording with more than one speaker.
Pitch (F0)	How high or low a voice sounds.
Metadata	Information that describes the speakers, recordings, and corpus, rather than the speech itself.

Overview

Language	Mandarin Chinese, including regional varieties
Speakers	54
Speech	≈10.5 hours
Sources	Public videos on Bilibili and YouTube
Speech genres	Interviews, podcasts, topical vlogs
Transcription	Automatic transcription (Whisper), reviewed by trained annotators
Time alignment	Word and phone level, produced automatically (Montreal Forced Aligner) and checked by hand, stored as Praat TextGrids
Identity metadata	Self-described only; separate fields for sex and gender identity, and sexual orientation
Access	Controlled, under the CQC Data Use Agreement
Documentation	Datasheet

Units

Speaker: an individual whose speech is represented in CQC, identified only by a CQC speaker ID.
Source recording: the original public video.
Session: the speech of one CQC speaker taken from one source recording. Genre is coded per session.
Segment: a stretch of a session that is transcribed and aligned.

A speaker may contribute several sessions. Report speakers, sessions, and segments separately in analyses.

Sources and inclusion

Each speaker passed every step below. To protect speakers, we do not publish the names of channels or videos.

Candidate discoverySearches on Bilibili and YouTube with Chinese and English terms for sexual and gender minority identities.
Self-description screeningKept only speakers who explicitly described themselves as members of sexual or gender minority communities in their own videos or public profile.
LanguageMandarin speech, including regional varieties; other Chinese languages such as Cantonese excluded.
Recording qualityAt most two people speaking; no persistent background music, canned laughter, heavy noise, or heavy overlap.
Speaker inclusionAt least five minutes of usable speech per speaker; only the target speaker’s speech is included.

Metadata and annotation

Gender identity, trans status, sex characteristics, and sexual orientation are different things, so each is recorded in its own field.

Annotation fields
Level	Field	Values
Speaker	Gender identity	woman, man, non-binary, other, not stated
Speaker	Trans status	trans, cis, not stated
Speaker	Sex characteristics	intersex, not stated
Speaker	Sexual orientation	gay, lesbian, bisexual, pansexual, asexual, heterosexual, other, not stated
Speaker	Mandarin variety	only if self-stated
Session	Speech genre	highly scripted, in-between, spontaneous
Session	Recording conditions	indoor / outdoor; number of speakers (1 or 2)
Segment	Transcription and alignment	Chinese characters; word and phone boundaries

Processing

Audio preparationRecordings converted to a single standard audio format (details in the datasheet).
Separating speakersIn two-speaker recordings, speaker diarization (pyannote.audio) separated the two voices; segments assigned to the target speaker were checked by hand.
TranscriptionSpeech transcribed automatically with Whisper, then reviewed by trained annotators.
Text clean-upTranscripts converted to simplified characters; punctuation removed.
Time alignmentForced alignment with the Montreal Forced Aligner marked the start and end of each word and phone, producing Praat TextGrids; the boundaries were then checked by hand.

Data quality

Step	What was checked	What to keep in mind
Speaker separation	Target-speaker segments checked by hand.	Very short turns may still be misassigned.
Transcription	All 5,183 segments reviewed by trained annotators; 64.2% of automatic transcripts were accepted without change.	Reviewed transcripts can still contain errors, especially in mixed Chinese–English speech.
Time alignment	Produced automatically, then checked by hand.	Checked boundaries can still contain small errors; check them if your analysis relies on precise timing.

Further details are in the datasheet. Processing scripts are available on request.

Statistics

Identity figures are based on how speakers described themselves, not on categories assigned by researchers.

Speech duration

5–10 min

48.1%

10–20 min

37.0%

20–30 min

11.1%

> 30 min

3.7%

Speech genre

spontaneous

11.1%

in-between

85.2%

highly scripted

3.7%

In-between: planned or topic-guided speech that was not read from a prepared script.

Sex, gender, and sexuality

Pie chart, sex and gender identities: intersex 1.9%, female (cis woman) 11.1%, male (cis man) 40.7%, trans man (FTM) 18.5%, trans woman (MTF) 25.9%, non-binary 1.9%.

Pie chart, sexual orientation: asexual 9.3%, pansexual 7.4%, bisexual 1.9%, lesbian 5.6%, gay 37.0%, heterosexual 7.4%, no information 31.5%.

Speaker embeddings

Box plot: for Male, Female, MTF and FTM, each speaker’s distance to the median of their category in the ECAPA embedding space — Dispersion in the ECAPA embedding space. Each point is one speaker’s distance to the median of their category; boxes show the distribution in each category. The spread within each category shows variation within groups.

Other figures

92.6% of speakers were recorded indoors.

To reduce the risk of identifying individual speakers, we do not publish detailed combinations of identity and Mandarin variety.

Limitations

CQC includes speakers who met the inclusion criteria and who published relevant content on two video platforms. Who ends up in the corpus is therefore shaped by platform use, by who chooses to speak publicly, by the search terms used, and by recording quality. Report sample composition, group sizes, and limitations, and interpret results with care.

Representation. Cis women, bisexual, non-binary, intersex, and older speakers are few.
Platforms. Bilibili and YouTube users are not representative of Mandarin speakers in general.
Publicness. CQC includes only people willing to speak publicly online about their identity.
Genre. Most speech is in-between; highly scripted and spontaneous speech are limited.
Geography. Speakers’ regional backgrounds are uneven and often not stated.
Self-presentation. Public videos involve deliberate self-presentation for an audience, and speakers may adjust what they say to platform rules.
Labels. Local identity labels do not always correspond to English categories; analytic categories simplify them.
Time. Online speech and identity vocabulary change; recordings reflect when they were made.
Recording conditions. Microphones, rooms, and compression vary, as in any collection of online media.

Releases

Material may be removed between releases at speakers’ request. Always cite the release you used.

Release	Date	Speakers	Hours	Changes
1.0	October 2026	54	≈10.5	First release