The training split contains approximately 5.15 million caption-aligned cues and 6,311 hours of audio after retaining segments between 1 and 8 seconds.

Audio is stored cue-by-cue as mono OGG/Opus inside Parquet shards. Source audio is sampled at 16 kHz, and the Hugging Face Audio feature exposes it at 16 kHz. Opus uses its standard internal 48 kHz representation when decoded before resampling. Text is stored as provided by the aligned WebVTT captions without additional normalization.

Fields:

id: Stable cue identifier.
audio: Embedded OGG/Opus cue audio.
text: Caption transcription.
caption_quality_score: Nullable float64 pilot ranking of caption–audio agreement. Higher scores indicate lower estimated wrong-caption risk, but are not calibrated probabilities. A null means no score was produced.
language: Automatically inferred spoken-language identifier from the cue or its programme (mostly ISO 639-1); und marks detected near-silence during per-cue classification.
duration: Cue duration in seconds.
source_id: Source programme identifier.
start: Cue start time in the source programme.
end: Cue end time in the source programme.

Configurations

The default configuration contains all 5,146,247 cues and retains the original caption, audio and metadata values. A quality score was produced for 3,218,383 cues; the other 1,927,864 have a null score and remain in default.

The danish-clean configuration contains 2,081,171 quality-filtered cues. It has the same fields and train split as default. Inclusion requires an automatically inferred language of da, a source_id not ending in .da_foreign, and a non-null caption_quality_score of at least 0.8616247194333528. This is a best-effort subset, not a manually verified Danish corpus.

Caption quality scoring

Danish-labelled cues outside .da_foreign tracks were assessed with Whisper tiny forced to Danish. Caption text was compared with its Whisper hypothesis using normalised character and word error rates, multilingual E5-small text embeddings, and programme-context similarity. A frozen five-feature pilot ranker combines these signals. The score is 1 - uncalibrated wrong-caption risk: it is useful for ranking, not a calibrated probability or a guarantee that a caption is correct. Whisper's own transcription errors can affect the score.

The danish-clean cutoff corresponds to estimated wrong-caption risk at most 0.1383752805666472. It prioritises rejecting unrelated or wrong captions while retaining related but imperfect captions. The cutoff was selected retrospectively after inspecting labelled test examples and has not been independently validated; corpus-wide precision or recall is unknown. The review labels and pilot model coefficients are not distributed with the dataset.

Quality scoring used source revision 9354bd9ded6605432e14e4f50bcb141213f48af8. The scorer implementation is recorded in syv-ai/hviske commit 079b784eb2216e90b7a6e2c4f8fd73b4d4e6f35b; initial single-server shards used e51728f862e8bf352476313596f01c8eecd801b2 with the same frozen features.

Streaming

from datasets import load_dataset

stream = load_dataset("syvai/drtv-asr", split="train", streaming=True)
example = next(iter(stream))

danish_clean_stream = load_dataset(
    "syvai/drtv-asr", "danish-clean", split="train", streaming=True
)
danish_clean_example = next(iter(danish_clean_stream))

Language identification

The language field was inferred from embedded audio using Whisper-small at model revision 973afd24965f72e36ca33b3055d56a652f456b4d. Up to 32 long cues sampled across each programme timeline were classified. If at least 90% agreed (stricter for some subtitle tracks), with no adjacent minority-language signal, the consensus language was assigned to all cues in that programme. Mixed and uncertain programmes were classified cue-by-cue. Audio and all other columns were preserved byte-for-byte at the value level. Language switches within otherwise uniform programmes, short cues and non-speech audio can still be mislabelled. Regional English accents may occasionally be mistaken for Welsh. These are automatic predictions, not manually verified annotations. The later quality-scoring pass added only caption_quality_score; pre-existing values remain unchanged.

Downloads last month: 97

Total file size:

98.4 GB