Time-frequency Dependent Multichannel Voice Activity Detection
Sebastian Stenzel, Jürgen Freudenberger · ITG Symposium of Speech Communication · 2014
This work proposes a method to determine voice activity for each time-frequency point in the noisy microphone signals of a microphone array. The speech signal as well as noise signals are assumed to be multivariate Gaussian random variables. Based on this signal model a generalized likelihood ratio test is derived. This likelihood ratio test can be simplified to a threshold test that compares the current a posteriori signal-to-noise ratio for each timefrequency point with a predetermined threshold. The theoretical results as well as the simulation results indicate that voice activity is well approximated.