SpeakerCard-1M (VoxCeleb release)

Evidence-grounded Speaker Card corpus for in-the-wild speaker verification.

SpeakerCard-1M is a speaker-centric resource built on VoxCeleb1/2 under a tool-first, LLM-last pipeline: ten acoustic probes extract field-level evidence, a schema separates relatively stable traits (gender, age band, accent, pitch band, timbre, language) from utterance-level states (emotion, channel, environment, speaking rate), and a constrained LLM verbalizes the structured fields into four bilingual (EN + ZH) caption styles per record. The release accompanies two cross-modal verification protocols:

• Bidirectional Speaker-Text Retrieval (T2S-R / S2T-R)
• Attribute-Conditioned Verification (AC-Verify) with two metrics:
  • CF — counterfactual rejection on style-symmetric LLM-rewritten distractors
  • Hard — near-miss rejection on mined hard-negative speaker cards

Scope note. This release covers the VoxCeleb portion only. The CN-Celeb portion described in the companion paper is not redistributed here; consult the CN-Celeb maintainers directly for that subset. See "License" below.

Release contents

Layer	File	Records
Speaker cards (EN + ZH renderings of VoxCeleb speakers)	data/speaker_cards/all.jsonl.gz	~40K
AC-Verify CF cards	data/ac_verify_cf/cf_cards.jsonl.gz	~156K
AC-Verify Hard triplets (train / val / test)	data/ac_verify_hard/{train,val,test}.jsonl.gz	per stats.json
Speaker-ID splits	splits/{train,val,test}_speakers.txt	text lists

Audio is not redistributed. Each record references VoxCeleb-relative speaker / utterance keys; obtain raw audio from the original VoxCeleb1/2 (Nagrani et al., Chung et al.) releases under their original license.

Field schemas

See schema.json for full per-record field documentation.

Data construction

Four-stage QC pipeline (malformed-output detection → MinHash near-duplicate filtering → identity-only state-leakage rejection → mDeBERTa-v3 NLI consistency check). LLM verbalizer: Qwen2.5-72B-Instruct (temperature 0.7, deterministic per-call seeds, three paraphrase variants per (speaker, language)). LLM counterfactual rewriter: Qwen3-32B-Instruct (temperature 0, schema-validated, minimal-edit prompt). Probes: language-ID (Whisper), gender/age (audEERING), accent (English classifier), timbre (voice-quality), emotion (SER), channel/environment (Brouhaha), pitch (three-system consensus over Praat / CREPE / RMVPE).

Statistics

See stats.json for per-layer record counts, language breakdown, and corpus-level distribution.

Splits

Train / val / test are speaker-ID disjoint at the VoxCeleb level. See splits/ for the canonical speaker lists.

Intended use

The release is intended for research on evidence-grounded speaker verification, cross-modal retrieval, and audio-language model evaluation. Released cards are not gold-standard trait labels; we retain per-field confidence and stability metadata so downstream users can filter low-confidence attributes.

Limitations

• Annotation quality is upper-bounded by the underlying probes; accent, timbre, and emotion remain noisier than gender / language-ID under in-the-wild conditions.
• Identity-overlap deduplication against external lists (SpeechBrain, VoxCeleb-Enhanced) is not applied; cross-VoxCeleb-ID identity overlap is possible.
• Hard-negative card difficulty is not calibrated against human judgments.
• A larger human-labeled pitch audit remains future work.

License

The cards, probe evidence, splits, and triplets are released under Creative Commons Attribution 4.0 International (CC BY 4.0), matching the upstream VoxCeleb1/2 metadata license. Source audio is not redistributed and retains its original VoxCeleb license; users must obtain audio directly from the VoxCeleb maintainers.

Dual-use note

Released as a queryable evidence artifact for interpretability and verification research, not as forensically valid speaker characterization. We do not endorse use of the corpus for identification, surveillance, or voice cloning.

Authors

Junyi Peng, Oldřich Plchot, Xiao Song, Dading Chong, Lichun Fan, Hang Su, Themos Stafylakis, Junjie Li, Kong Aik Lee, Shuai Wang, Jian Luan, Jan Černocký.

Affiliations: Brno University of Technology · Peking University · Xiaomi · Athens University of Economics and Business · The Hong Kong Polytechnic University · Nanjing University.

Citation

@misc{peng2026speakercard,
  title  = {SpeakerCard-1M: An Evidence-Grounded Speaker Card Corpus for In-the-Wild Speaker Verification},
  author = {Peng, Junyi and Plchot, Old\v{r}ich and Song, Xiao and Chong, Dading and Fan, Lichun and Su, Hang and Stafylakis, Themos and Li, Junjie and Lee, Kong Aik and Wang, Shuai and Luan, Jian and \v{C}ernock\'{y}, Jan},
  year   = {2026}
}