Algorithm of Spontaneous Speech Recognition Based on A^<**> Search

自发语音识别被认为是一个图搜索问题，它考虑了声学模型、词典、语言模型等各种约束条件。为了在不降低识别性能的前提下减少识别处理的计算量，对自发语音识别的一些关键技术进行了研究。(1)声学模型和说话人自适应利用有限的训练数据集进行上下文相关声学建模的重要方面是如何绑定模型参数和如何处理看不见的上下文。提出了一种基于决策树的连续状态分裂算法，并证明了该算法生成的HM网具有较高的准确率，能够表示任意上下文。研究了基于MAP估计方法的声学模型参数说话人自适应。(2)快速匹配和似然归一化在大词汇量单词识别中，研究了一种快速预选候选单词的方法。对输入语音进行音素识别，并从输入语音中获得最佳音素序列。为了选择候选单词，使用最优音素序列执行DP匹配。通过输入语音和基于HMM的单词模型之间的维特比评分来验证候选单词。从EDR语料库、500万词日语语料库出发，构建了语言模型和任务适应N元语法语言模型。在不同的训练文本大小、词汇量和截止条件下对模型进行了研究。实验结果说明了在一定的训练文本大小下的最优条件。我们进行了另一项关于任务适应的实验。将对话中的N元语法模型与EDR语料库中的N元语法模型混合，减少了约60%的困惑。

Spontaneous speech recognition is regarded as a problem of graph search considering various restrictions through acoustic model, lexicon, language model and so on. In order to reduce a computation amount for recognition processing without degradation of recognition performance, some key technologies of spontaneous speech recognition were investigated.(1) Acoustic model and speaker adaptationThe important aspects of context-dependent acoustic modeling using a limited training data set are how to tie the model parameters and how to handle the unseen contexts. We proposed the decision tree-based successive state splitting algorithm, and showed that HM-Net generated with this algorithm had high accuracy and enabled to represent any contexts. Speaker adaptation of acoustic model parameters based on MAP estimation method was also investigated.(2) Fast matching and likelihood normalizationIn large vocabulary word recognition, a fast preselection of word candidates was investigated. Phoneme recognition of input speech was carried out and an optimal phoneme sequence was obtained from the input speech. To select word candidates, DP matching was executed with the optimal phoneme sequence. The word candidates were verified by Viterbi scoring between input speech and HMM-based word model. Normalization technique of word likelihood for spontaneous speech recognition was also investigated.(3) Language model and task adaptationN-gram language models were constructed from EDR corpus, 5-million-word Japanese corpus. The models were investigated under various conditions about training text size, vocabulary and cutoff condition. The result of experiments clarified the optimum condition under a certain training text size. We carried out another experiments about task adaptation. An N-gram model from a dialog was mixed with the N-gram from EDR corpus, which made about 60% reduction of perplexity.