Detecting pronunciation erroneous tendency (PET) can provide detailed instructive feedback for second language learners in computer aided pronunciation training (CAPT). In this paper, we proposed to apply soft targets from various models to improve the detection performance of PET. First, we examined the effectiveness of soft targets in three single systems by replacing hard targets with soft targets directly for mispronunciation detection. Furthermore, we proposed two kinds of methods using multi-model soft targets in this paper: 1) explicit combination, which used multi-model soft targets as the final targets by weighted linear combination; 2) implicit combination, which employed the multi-task framework to combine soft targets. Experimental results showed that the detection performance of PET could be improved by using both single soft targets and multi-model soft targets. Moreover, using multi-model soft targets within multi-task framework achieved the best results in pronunciation error detection task, and it was more efficient than conventional ensemble methods which required multiple decoding runs or forward passes.

检测发音错误倾向（PET）可以在计算机辅助发音训练（CAPT）中为第二语言学习者提供详细的指导性反馈。在本文中，我们建议应用来自各种模型的软目标来提高PET的检测性能。首先，我们通过用软目标直接替换硬目标以进行错误发音检测，在三个单一系统中检查了软目标的有效性。此外，本文提出了两种使用多模型软目标的方法：1）显式组合，通过加权线性组合将多模型软目标作为最终目标。2）隐式组合，它采用多任务框架来组合软目标。实验结果表明，同时使用单个软目标和多模型软目标可以提高PET的检测性能。此外，在多任务框架中使用多模型软目标在语音错误检测任务中获得了最佳结果，并且比需要多次解码运行或前向传递的传统合奏方法更加有效。