Non-stationary noise model compensation in voice activity detection

This paper proposes methods for acoustic pattern recognition in dynamically changing noise. Parallel model combination and vector Taylor series model compensation techniques are used to adapt acoustic models to noisy conditions are applied together with a time-varying noise estimation algorithm. The noise estimation produces biased noise estimates and therefore we propose methods to accommodate the compensation to the bias. We apply the methods in robust voice activity detection, where frame-wise speech/non-speech classifier is first trained in clean conditions and then tested in and adapted to non-stationary noise conditions. The simulations show that a model compensation with the time-varying noise estimator improves clearly the accuracy of voice activity detection.