lilgoose777/kantipur-interview-data3

Nepali Speech Dataset (YouTube-sourced)

83 labeled speech segments, split by channel (not by individual video) so the same speaker/recording can't appear in more than one split.

Splits

• train: 83 segments
• validation: 0 segments
• test: 0 segments

Transcript columns — read this before training

Each segment carries three transcript variants. They are NOT interchangeable:

• text_original — the YouTube caption text (if any) that overlapped this segment's time range, in whatever spelling/style convention that caption used.
• text_whisper — this project's fine-tuned Whisper model's own transcription of the segment's audio.
• text / text_final — the transcript actually selected as this segment's label (see transcript_source for which one won). This is a mix of both conventions across the dataset — some segments carry caption-style text, others carry Whisper-style text — so treat text_final as a single working label, not a normalized ground truth, and check transcript_source/label_tier per segment if your use case needs one consistent convention.

Label trust tiers (label_tier)

transcript_source is collapsed into a coarser trust label:

• gold (youtube_manual_caption) — human-authored caption.
• silver (youtube_auto_caption) — YouTube's own ASR caption.
• bronze (custom_whisper) — this project's Whisper model's own output.

Silver and bronze are both machine-generated transcripts, and the Whisper model used for bronze was fine-tuned on a corpus with its own systematic error patterns — treat bronze as a separate, lower-confidence pool rather than blending it silently with gold for anything where transcript accuracy matters most.

• gold: 11 segments (13.3%)
• bronze: 72 segments (86.7%)

Quality tiers (quality_tier)

A/B/C reflect a combination of audio cleanliness (clipping, VAD speech coverage, estimated SNR), forced-alignment confidence, and language-ID confidence. These thresholds are heuristic starting points, not independently calibrated values — if precision matters for your use case, spot-check a sample against human judgment before trusting a tier at face value.

music_likelihood and overlapping_speech are best-effort heuristic flags (not trained classifiers) that cap a segment at tier B rather than rejecting it outright — a false positive should cost a tier, not the clip.

Content mix

• interview: 100.0%

KL divergence from the target content mix: 1.3863 (0 = matches target exactly).