Real-time wind buffet noise detection

Wind buffet noise in a microphone signal is detected using a per-frequency probability of speech estimate as well as short-term and long-term low-frequency energy. Using the probability of speech presence estimate, the buffet-no-speech condition can be accurately detected. But the probability of speech presence, by itself, is insufficient for distinguishing between the buffet-speech and either of the no-buffet conditions. It can be assumed that, if wind buffeting is occurring, it is occurring during both speech and non-speech segments to help distinguish between the other possible states. That is, the probability difference may be used as the criteria for entering the buffet-no-speech state, and then some other information (e.g., low frequency energy) may be used to determine when to transition to the buffet-speech state or one of the no-buffet states once the probability difference criteria is no longer met. 实时风冲击噪声检测。使用每频率语音的概率估计以及短期和长期的低频率能量来检测麦克风信号中的风冲击噪声。使用语音存在的概率估计可以准确地检测冲击无语音条件。但语音存在的概率单独地不足以在冲击语音与无冲击条