VAD

Why we included this project

Voice activity detection, figuring out where speech actually starts and stops in an audio stream, is a small piece that can quietly make or break a speech pipeline. This repository from KAIST bundles four VAD classifiers: a plain DNN, a boosted variant, an LSTM recurrent network, and an adaptive context attention model, all fed with multi-resolution cochleagram features rather than ordinary spectrograms. It also ships a manually annotated dataset recorded on a phone across four noisy environments, useful both for benchmarking and for seeing what realistic training data looks like. The honest caveats are that feature extraction depends on MATLAB and runs slowly, and the training code targets TensorFlow 1.x, so treat this as a reference for understanding the full workflow rather than something to drop straight into production. Students walking through how these models are trained end to end, or a team prototyping a segmenter before moving to a maintained library, will get the most out of it.