Real-time wind buffet noise detection

The probability of speech being present in a microphone signal (112) subject to wind buffet noise (eg. in a vehicle) is estimated by starting a state classifier 200 in an assumed no buffeting, no speech state and transitioning to a buffet noise, no speech state when the difference between the average speech presence probabilities in two frequency bands (lower and higher) exceeds a threshold. When this threshold is no longer exceeded, the state is further transitioned to speech + buffet, speech without buffet or no speech no buffet depending on the difference between short-term low frequency energy and the long-term energy floor.