Stability of a voice activity detector based on source separation

Source separation techniques been applied to the problem of voice activity detection (VAD) and a method has been developed which is able to perform well in very adverse. Environments (-5 to -10 dB). The optimisation procedure is totally iterative and therefore suitable for use in real-time, dynamically adapting situations. Stability considerations are addressed in this paper and simulation results are given. These show the new technique to be very robust to high levels of additive noise, with a relatively low complexity.