Speaker Conditional Sinc-Extractor for Personal VAD

This study explores Sinc-convolution's novel application in Personal Voice Activity Detection (PVAD).The Sinc-Extractor (SE) network, developed for PVAD, learns cutoff frequencies and band gains of sinc functions to extract acoustic features.Additionally, the speaker conditional SE (SCSE) module incorporates speaker information from high-dimensional d-vectors into low-dimensional acoustic features.SE-PVAD and Vanilla PVAD have similar model size and computing load, while SCSE-PVAD is more compact with shorter inference time as it excludes speaker embedding.Evaluated with concatenated utterances from the LibriSpeech corpus, SE-PVAD outperforms Vanilla PVAD significantly.SCSE-PVAD matches Vanilla PVAD's performance but reduces input feature dimensionality and network complexity.Thus, SCSE-PVAD can function like a typical VAD, accepting only acoustic features, making it suitable for low-resource wearable devices.