Blaze-2/khmer-podcasts-asr

Khmer Podcasts ASR — pseudo-labelled

VAD-segmented audio from blaze-voice-ai/khmer-podcasts with machine-generated transcripts.

Contents

Segments	157,590
Speech	343.1 h (from 383.1 h of source audio; 90 % speech after VAD)
Audio	24 kHz mono, cut losslessly from the source clips (no resampling)
Segment length	0.3 – 20 s (median 6.1 s) — matches the model's training filter
Source clips	1,359 (of 1,366; 7 were digitally silent upstream and produced no segments)
Channels	VTV5 – Nhịp sống đồng bào · Cần Thơ – Tiếng Dân tộc

Columns

• audio — 24 kHz mono segment
• text — pseudo-label transcript
• duration, start, end — seconds; start/end are offsets into the source clip
• video_id, channel — provenance, joinable back to the source dataset
• chars_per_sec, is_empty, has_space, bad_density, is_clean — quality flags

Quality flags

Rows are not filtered — flags are provided so you choose your own threshold.

Flag	Rows	Share	Meaning
is_empty	1,583	1.0 %	Model emitted nothing
has_space	12,201	7.7 %	Khmer has no inter-word spaces; the model card names emitted spaces as its hallucination signature
bad_density	~3,549	2.2 %	Character rate outside 3–25 chars/s (jingles, music, missed speech)
is_clean	~140,000	~89 %	None of the above

ds.filter(lambda r: r["is_clean"]) gives roughly 305 h.

Regenerating

start/end are exact offsets into the source 24 kHz clips, so every segment can be recut deterministically from blaze-voice-ai/khmer-podcasts without rerunning VAD.

⚠️ Copyright

Inherited from the source dataset: these are third-party broadcast recordings (VTV5 / Can Tho TV) collected from YouTube. Rights remain with the original broadcasters. Provided for research use. Verify you have the necessary permissions before redistribution or commercial use.