JackyHoCL/legco-yue-official-transcripts

LegCo Yue Official Transcripts - ALL VENUES (ASR training set)

119204 Cantonese speech chunks (~914 h) from all Hong Kong Legislative Council meetings with official AI realtime transcripts: plenary (Chamber) and committee/panel meetings (2025-07 to 2026-10, 396 meetings). Audio: official webcast Floor recordings, 16 kHz mono PCM16 WAV, embedded in parquet (Common-Voice style). Transcript: official AI captions (traditional Chinese, written form), chunked 2-30 s.

Files

• 395 parquet shards: {train,validation,test}-{i}-of-{n}.parquet (~256 MB each); splits are disjoint by meeting
• metadata.json - per-meeting provenance & stats

Usage

from datasets import load_dataset
ds = load_dataset("JackyHoCL/legco-yue-official-transcripts", split="train")
ex = ds[0]
import soundfile as sf, io
arr, sr = sf.read(io.BytesIO(ex["audio"]["bytes"]))  # 16 kHz mono
ex["sentence"]     # written-form Cantonese (traditional)
ex["venue"]        # "Chamber" | "Conference Room 1" | ...

Schema

audio (Audio, struct {bytes, path}), sentence, speaker, speakers, start_s, end_s, duration_s (absolute offsets within the meeting), meeting_id, meeting_date, venue, chunk_idx, id

Notes

• All-venues release (replaces the earlier plenary-only 107 h version of this repo): same rows for the shared 12 plenary meetings + 384 committee meetings.
• Labels are AI realtime captions (僅供參考): written form, not phonetic; fine for ASR pretraining/fine-tuning, not a verbatim Hansard substitute.
• Speaker attribution is heuristic (caption speaker-change cues); multi/unknown labels expected ~5-15%.
• Committee audio is from smaller conference rooms (closer mics, different reverberation) than plenary.