DGC-AI/dgc-meeting-asr

DGC meeting ASR (Khmer, pseudo-labelled)

Internal DGC use only. Built from the internal DGC-AI/dgc-meeting-recordings collection.

Contents

Recordings	841 meetings
Segments	149,106
Speech	476.0 h (silence removed)
Split	train only (no held-out split)
Shards	segmented/train/vm0-*.parquet (398) + vm1-*.parquet (393), ~52 GB

Columns (same layout as the COHP segmented sets):

• audio {bytes, path}: 16 kHz mono WAV
• transcription: Khmer text
• duration: seconds
• confidence: mean token log-prob from the decoder (higher = more confident)
• recording_id
• start_sec, end_sec: position in the original recording

How it was made (2026-10-15/16)

• Decode: each recording decoded with PyAV, then Silero VAD (faster-whisper's VAD, 500 ms min silence, 200 ms padding).
• Segmentation: speech merged into segments of 25 s or less, splitting where the gap is over 1 s. Segments under 1 s dropped.
• Transcription: Whisper large-v3-turbo Khmer R8 (DGC-AI/whisper-large-v3-turbo-khmer-r8), CTranslate2 int8 on CPU, greedy, no previous-text conditioning.
• Where it ran: two DGC data-centre VMs, each taking every second recording sorted by size (meet_full.py K k).

Caveats

• The labels are machine pseudo-labels, not human transcripts. R8 scores 10.6 % CER on DGC-new, and meeting audio (far-field, overlapping speakers, code-switching) will be worse.
• Recommended use: filter by confidence and/or agreement with a second model (e.g. Qwen3-ASR Q7T) before training.
• Never use as an evaluation set.
• Overlapping speech is transcribed as one stream. There are no speaker labels.
• Number style follows R8 (digits).