基于束搜索法的基音标注新方法

A New Method of Pitch Marking Based on Beam Search Algorithm

  • 摘要: 基音标注在语音合成等方面起着重要作用。目前使用比较广泛的动态规划基音标注算法,约束准则大都比较简单,采用的动态规划算法往往偏重于局部最优,而非全局最优。基于此,提出了一种基于束搜索法的基音标注新方法。除周期与幅度外,引入了图形与位置作为约束准则,更严格地筛选基音标注候选点,并采用束搜索的方式,从全局最优出发,兼顾局部最优,进行基音标注。此外,为了提高基音估计准确度,获得更精准的基音标注,还提出了一种基于束搜索法的基音轨迹提取方法,从多种不同的基音检测算法结果中提取基音轨迹。仿真结果表明,与传统的动态规划基音标注算法相比,基于束搜索法的基音标注新方法具有更高的准确率。仿真中,新方法的平均准确率为98.57%,而传统方法的平均准确率为94.70%。

     

    Abstract: Pitch marking plays a key role in speech synthesis and other areas. The pitch marking algorithm based on dynamic programming, which is commonly used at present, has some shortages. For instance, its constraint criterions are usually too simple, and the dynamic programming method it adopts considers local optimum heavily but global optimum. On account of this, a new method of pitch marking based on beam search algorithm is proposed. Except pitch period and amplitude, figure and index position are brought in the constraint criterions to select the candidates more restrictedly. Further more, beam search algorithm is employed to extract the pitch marks, for basic consideration of the global optimum and giving due consideration to the local optimum. Besides, a pitch contour extraction method based on beam search algorithm is proposed, which extracts the pitch contour from the outputs of several pitch detection algorithms, to obtain more accurate pitch marking result by improving the precision of pitch period estimation. Simulation results indicate that, compared to the traditional pitch marking algorithm based on dynamic programming, the new method of pitch marking based on beam search algorithm owns higher accuracy rate. In the simulation test, the average accuracy rate of the new method is 98.57%, while the average accuracy rate of the traditional algorithm is 94.70%.

     

/

返回文章
返回