This release contains 189,859 automatically transcribed, time-bounded utterances consolidated across 174 Brisbane City Council meeting recordings (707.5 hours of time-bounded audio).

Each utterance row contains the verbatim transcript text, word-level timing, agenda-section metadata recovered from the official minutes, canonical speaker attribution, dated official metadata and remuneration, recording provenance, and source start/end times. Adjacent source units are consolidated only for the same resolved speaker and never across conflicting official turns. Unresolved speaker labels are not merged.

The default configuration embeds 16 kHz, 32 kbps mono Opus audio in Parquet for direct in-browser streaming and playback in the Hugging Face Dataset Viewer.

The companion audit configuration preserves source speaker labels, assignment methods, diarization clusters, official-turn links, official text, alignment methods, model identifiers, and metadata source identifiers. The speakers.parquet table reports all dated roles, wards, and parties observed for each resolved speaker rather than presenting the first corpus value as static. meetings.parquet provides the recording-level inventory. recordings.parquet provides technical audio metadata for full meeting recordings.

Limitations

Transcripts and speaker attribution are automatic. Official minutes are edited records and not verbatim references. Sex is taken from public records where explicit and otherwise inferred from the speaker's title or given name. Age is left empty unless a suitable public source supports it. Ages calculated from a year-only record, a reported age, or an explicit age-linked milestone are marked approximate and may differ by one year. Approximate birth years are not published as factual birth years. Earlier unsupported birth years and interest summaries were removed. Users should validate critical examples against the source audio.

Agenda-section metadata

agenda_section is a canonical label derived from explicit official-minutes headings; agenda_heading_raw preserves the heading text. Direct official-turn links have high confidence. Unlinked utterances are assigned only when high-confidence anchors on both sides have the same heading within five minutes; these rows have medium confidence. Other rows remain unassigned. Section labels describe agenda structure and do not assert whether speech was prepared, spontaneous, adversarial, or otherwise interactional.

Downloads last month: 241

Total file size:

22 GB