Semantic-VAD EOT

End-of-turn (semantic VAD) turns built from word-level forced alignments, schema-compatible with livekit/eot-bench-data.

Each row is one user turn: an audio clip (16 kHz mp3), its words, and ordered silence_spans. Per the eot-bench convention the last silence span is the true end-of-turn (eot); earlier spans are mid-turn hold pauses (labels positional, not stored).

Splits

For every data type, all shards except the last form the train base; that type's last shard is split 50/25/25 into extra train / validation / test (shuffled, seed 42).

Configs (viewer dropdown)

• all (default) — every language + Malaysian subset.
• 13 languages: en, fr, de, it, ja, ko, zh, pl, pt, ru, es, th, tr.
• 5 Malaysian subsets (word-level, no synthetic): ms_dialects, ms_imda, ms_malaysian, ms_parliament, ms_science_english.

Sources: AAdonis/multilingual_audio_alignments, malaysia-ai/Malaysian-STT.

from datasets import load_dataset, Audio
ds = load_dataset("Scicom-intl/semantic-vad-eot", "en")   # ('train', 'validation', 'test')
te = load_dataset("Scicom-intl/semantic-vad-eot", "ms_imda", split="test")
ds = ds.cast_column("audio", Audio(decode=False))