Robust speech understanding against inter-speaker variation and ungrammatical utterances based on high-accuracy speech recognition and semantic driven parsing method

本研究的目的是构建一个稳健的语音理解系统，以对抗说话人之间的变异和非语法话语。为了实现这种稳健的系统，我们开发了一种结合说话人自适应方法和语义驱动分析方法的高精度语音识别器。1.HMM音素识别器的说话人自适应我们提出了一种针对说话人之间变化的连续语音输入的说话人自适应方法。我们将最大后验概率估计应用到基于Pair-Wise Bays分类器的连续密度隐马尔可夫模型(HMM)中，作为语音分类器。我们对8名说话人进行了顺应性实验评估。结果表明，说话人自适应模型的关键词识别率达到了80.2%，比基线模型提高了11.0%，而对另一个说话人的识别率则有所下降。2.词/短语检测方法即使在自发语音中，大多数单词和短语也是正确发音的。然后，我们需要一种词/短语识别方法来进行健壮的句法分析。为了提高这些Spotter的准确率，我们开发了一个启发式语言模型来模拟目标单词/短语的其余部分。此外，我们还实现了一个孤岛驱动的PRASER，它可以跳过填充的停顿和未登录的单词。现有的方法很难处理这些现象，因为它的前提是完整的句子输入。对于这些零碎的话语，我们尝试使用带有松弛的增量式句法分析方法。在实现上，我们使用标记传送器将输入片段集成到识别的计划结构中。

The aim of this research is to construct a robust speech understanding system against inter-speaker variation and ungrammatical utterances. In order to implement such robust system, we develop a high accuracy speech recognizer with a speaker adaptation method and a semantic driven parsing method.1.Speaker adaptation of HMM phoneme recognizerWe develop a speaker adaptation method using continuous speech input against inter-speaker variation. We use maximum a posteriori probability estimation to Continuous density Hidden Markov Model (HMM) based on Pair-Wise Bays Classifiers as the phone classifier. We performed experimental evaluation of adaptation to 8 speakers. As a result, the keywords recognition rate of the adapted model of a speaker reached 80.2 %, Which is higher by 11.0 % than that of the baseline model, while the accuracy is lowered for another speaker.2.Word/Phrase spotting methodEven in spontaneous speech, most words and phrases are correctly uttered. Then, we need a word/phrase spotting method for robust parsing. In order to increase accuracy of these spotter, we develop a heuristic language model that models the rest of target word/phrase. Also, we implement a island-driven praser that can skip filled pauses and unknown words.Robust speech parser by incremental analysisIn natural dialogues, fragmentary utterances are frequently used. Existing approach can hardly deal with these phenomena because it presupposes a complete sentence input. We try to use incremental parsing method with relaxation to such fragmentary utterances. For implementation, we use marker passer to integrate input fragment to recognized plan structure.