基于图检测方法的语音信号检测

Speech Signal Detection Based on Graph Detection Method

大部分检测方法针对平稳信号设计,使用单一统计模型分析检测性能,面对语音信号及复杂环境噪声时,效果常因模型失配而变差。为探索具有较高鲁棒性的语音信号检测方法,将语音信号与不同环境噪声混合并将其映射为图,设计基于图结构的语音信号检测方法。通过实验比较该方法与传统M2M4方法的性能。对比受试者工作特征(ReceiverOperating Characteristic,ROC)曲线和曲线下面积(Area Under Curve,AUC)值,图检测方法均优于M2M4方法的检测性能。在不同实验条件下,基于图检测方法的信号检测具有更优的检测性能。

Most detection methods are designed for stationary signals and use a single statistical model to analyze the detection performance.Therefore,when facing speech signals and complex environmental noises,their effectiveness often deteriorates due to model mismatch.In order to explore a speech signal detection method with high robustness,a graph-based speech signal detection method is designed by mixing speech signals with different environmental noises and mapping them into graphs.The performance of this method is compared with the traditional M2M4 method through experiments.Comparing the Receiver Operating Characteristic(ROC)curves and Area Under Curve(AUC)values,the graph detection method outperforms the M2M4 method.Under different experimental conditions,the graph-based signal detection method has better detection performance.