中国科学院自动化研究所机构知识库

Voice activity detection (VAD) is an important pre-
processing part of many speech applications. Context information
is important for VAD. Time-delay neural networks (TDNNs)
capture long context information with a few parameters. This
paper investigates a TDNN based VAD framework. A simple
chunk based decision method is proposed to smooth raw pos-
teriors and decide border points of utterances. To evaluate
decision performance, a metric intersection-over-union (IoU) is
introduced from image object detection. The experiment results
are evaluated on Wall Street Journal (WSJ0) corpus. Frame
classification performance is measured by area under the curve
(AUC) and equal error rate (EER). Compared with long short-
term memory baseline, the TDNN based system achieves a
41.26% EER relative reduction on average in matched noise
condition, and relative improvement of average AUC is 3.82%.
Proposed decision method achieves an 18.74% IoU relative
improvement on average compared with moving average method
on average.