非凸段长分布隐含Markov模型的搜索算法

基于段长分布的隐含Markov模型(DDBHMM)可解决经典隐含Markov模型(HMM)的状态段长指数分布的问题,实现了基于凸性假设的搜索。为解决非齐次模型的搜索算法问题,提出采用混合Gauss分布来拟合非凸段长分布,用子状态拆分的方法来实现非凸段长分布DDBHMM识别算法。在音乐信号识别上的实验表明:该方法在召回率提高1.1%的情况下,使准确率提高约10%。该方法实现了非凸段长分布HMM的识别算法,并且对于其他非凸段长信号具有推广价值。

Hidden Markov model search algorithm for non-convex duration distributions

A duration distribution-based hidden Markov model (DDBHMM) was developed to replace the exponential duration distribution limitations of classical hidden Markov model (HMM) for non-convex signals. The search algorithm initially developed for convex duration distributions actually identifies non-convex duration distributions. The gaussian mixture density (GMD) algorithm is used to simulate the non-convex duration distribution, with the search algorithm using sub-state combination. Music signal recognition tests show a 10% percent precision improvement with (1.1%) recall ratio improvement. Therefore, this is an effective model and search algorithm for non-convex duration distribution HMM which can be extended to other non-convex signals.