语谱图共振峰的自动跟踪算法

本文介绍一个共振峰轨迹的自动跟踪算法，其特点是不借助于其它的信息来源，仅仅是基于语谱图信息，来确定语谱图上前四个共振峰频率的位置和它们关于时间轴的轨迹。算法由三个层面构成；第一层面是进行频率分布的分析，以决定一个最佳的共振峰搜索起始位置；第二层面是采用双搜索算法，来跟踪随时间变化的共振峰轨迹；第三层面是解决某些冲突现象，在共振峰跟踪问题上的难点之一是处理多个共振峰的合并现象和冗余峰值现象。

An algorithm for formant tracking on speech spectrograms

This paper describes a formant tracking algorithm which evaluates the frequency tracks of the four lowest formants. The algorithm takes segmented speech spectrograms of continuous speech as its input. There are three levels in this algorithm: The first level includes format frequency analysis technique to locate a best beginning position for formant tracking. The second level includes a method of formant tracking using two direction searching. The last level is a post-processing to handle some particualar events, such as merge of two peaks, or spurious peaks. An evaluation based on a large collection of recorded spectrograms is taken to test its performance, which compares the results computed by the algorithm with the results marked by a phonetics expert.