Voice Activity Detection (VAD) Test Dataset
This dataset is based on the test splits found in
multilingual_librispeech
dataset. It includes two binary labels:
speech: Indicates presence of speech ([0, 1]), computed using a dynamic threshold method with background noise estimation and smoothing.
confidence: A post-processing flag to optionally correct transient dropouts in speech. It is set to 1 by default, but switches to 0 for up to ~0.1 seconds (3 chunks of audio) following a transition… See the full description on the dataset page:
https://huggingface.co/datasets/guynich/multilingual_librispeech_test_vad.
