Method of detecting voice activity starts with by generating probabilistic models that respectively model features of speech dynamically over time. Probabilistic models may model each feature dependent on a past feature and a current state. Features of speech may include a nonstationary signal presence feature, a periodicity feature, and a sparsity feature. Noise suppressor may then perform noise suppression on an acoustic signal to generate a nonstationary signal presence signal and a noise suppressed acoustic signal. An LPC module may then perform residual analysis on the noise suppressed data signal to generate a periodicity signal and a sparsity signal. Inference generator receives the probabilistic models and receives, in real-time, nonstationary signal presence signal, periodicity signal, and sparsity signal. Inference generator may then generate in real time an estimate of voice activity based on the probabilistic models, nonstationary signal presence signal, periodicity signal, and sparsity signal. Other embodiments are also described.