一种低虚警概率的啸叫检测方法

Low false alarm probability howling detection method

啸叫检测是基于陷波器的啸叫抑制算法的关键.而传统的啸叫检测方法普遍存在虚警率偏高,易造成输出语音失真的问题.为此,提出了一种基于长时信号特定频带变化率的啸叫检测方法.首先,通过提取功率谱峰值,确定候选啸叫频点;然后,将候选啸叫频点作为中心频点,并在其两侧分别选取若干个频点组成目标特定频带;最后,测度长时信号在目标特定频带的变化率,并进行阈值判决.仿真实验结果表明,与传统的啸叫检测方法相比,文中方法在取得较高啸叫检出率的同时,可以更有效地控制虚警率.

Howling detection is the key of the notch-filter-based howling suppression algorithm. However, the traditional howling detection method with a problem of high probability of false alarm will lead to speech distortion easily. So a novel howling detection method based on long-term signal variability at the specific frequency band is proposed in this paper. First, we pick the candidate howling competent by means of extracting the power spectrum peak; then, we select several bins on both sides of the candidate howling competent as the target specific frequency band; lastly, whether howling occurs or not depends on the value of long-term signal variability at the target spectral frequency band. Simulation results show that compared with the traditional howling detection method, the probability of false alarm can be controlled more effectively while keeping a high probability of detection by the proposed method.