英语语调评测中韵律段切分算法

在英语考试评测的复杂噪声环境中,韵律段的选取面临着很大的困难。将韵律段的选取目标明确为寻找韵律核心段,并进一步细化为2步,即基于能量、自相关的清浊音切分与基于音素HMM模型的发音网络强制匹配结果的结合。其中,发音网络强制匹配充分利用了先验信息,而根据清浊音切分的向前向后延展是为了控制段长和舍弃非韵律核心段。新算法经3个库的测试,语调评测错误率平均下降了3．8％。

Unvocied-vocied Decision Research for English Intonation Assessment

Great difficulties confront the first step which selects the intonation segment in the complex noise environment of English test. So the segment goal is selected to find the kernel intonation segment. It consists of two steps： unvoiced-voiced decision based on energy and auto-correlation function, and combination with the force align result of pronunciation net based on phone-level HMM model. The net contains enough a prior information while extending forward and backward is to control the length of segment and to drop weak voiced ones. The new algorithm has been tested by three corpus. The total error rate of intonation assessment declined obviously.