A Study on Ambiguous Utterance Understanding for Speech Input

本文提出了一种基于连续参数隐马尔可夫模型（HMM）的最大后验概率估计（MAPE）的无监督连续拼接训练方法。识别器采用事先建立的非特定人模型，自动生成标签序列。在连续语音识别上的实验结果表明，该模型的识别性能与监督自适应模型相当。其次，提出了一种处理感叹词和未登录词的方法，使语音识别系统能够处理对话中的自发语音。我们使用测试句子集（包括感叹词和未登录词）对语音识别系统的性能进行了评估，证实了所提方法的有效性。第三，我们研究了能够理解所有用户输入的菜单引导的口语自然语言理解系统。这项工作的动机是以下事实，即用户无法理解说什么或如何说的计算机在自然语言。系统显示一个菜单，该菜单由可接受的内容词组成，用户从菜单中选择一个词，并说出包括该词的短语。实验结果表明，该系统对于新手用户来说，性能良好。

We proposed an unsupervised speaker adaptation method on sequencial concatenation training that used the theory of MAPE(Maximum A Posteriori probabitity Estimation) for continuous parameter HMM.In this method, we should only specify the syllable label sequence for the utterrance. The label sequences were provided automatically by the recognizer which used a speaker-independent model in advance. The experimental results on continuous speech recognition showed that the better model gave a performance comparable to that of supervised adaptation.Secondly, we proposed a method to process interjection and unknown words so that a speech recognition system could deal with spontaneous speech in dialog. We have evaluated the peerformance of our speech recognition system using test sentence sets including interjection or unknown words, and confirmed that the proposed method worked well.Thirdly we investigated the menu-guided spoken natural language understanding system that could understand all user's inputs. This work was motivated by the following fact that a user could not understand what to say or how to say to a computer in natural language. The system displays a menu that consists of acceptable content words and the usur chooses one word from the menu and speaks out phrase that includes the word. The experimental showed that our system performed well for the novice users.