Speaker-Independent Word Recognition Based on Recognition of Phonemes and Adaptation to Speakers

本课题取得的研究成果概括如下：1.研究成果。音素的识别我们的研究侧重于从音素在结构分析中起重要作用的角度来提高音素的识别率，即验证单词或闭合为词尾。结果，特别是在识别元音、元音、鼻音和塞音方面，取得了一些显著的进步。对于元音，我们开发了一种新的基于元音二元关系模型的识别技术。单词识别我们已经开发出一种强大的单词识别技术，它由两部分组成。第一部分是我们在字符识别研究中发展起来的一种模式匹配方法，称为联想模式匹配方法，在单词的粗分类中达到了很高的准确率。第二部分是结构分析，即基于音位识别的词语变异。适应说话人我们发现了在特征空间中表示的日语元音之间的二元关系，这种关系可以被视为几乎独立于说话人，并发展了上述元音识别技术。在对候选词语进行验证的过程中，使用元音识别技术可以实现对说话人的顺应。识别实验对10名MAL说话人的212个单词进行了单词识别，获得了比以往任何时候都高的识别率，即第一个条件数据的识别率为98.2%，第二个条件数据的识别率为99.6%。然而，这些结果是通过使用基于文献1的元音识别的结构分析和D.P.的方法获得的，因为我们的技术还不能在结构分析中使用音素识别。这是一个功能问题。

The results obtained in the research project are summarized as follows.1. Recognition of phonemesOur research stressed on improving the recognition rate of phonemes from the viwe-point that they play an important role in the structural analysis,i.e., verificatgion of words or closes as chandidates. As a result some significant improvements have been obtained especially in recognition of vowels, vowels, nasals and plosives. As for vowels we have developed a new recognition technique based on a binary relational model of vowels.2. Word recognition We have daveloped a powerful technique for word recognition which consists of two parts. The first part is concerned with such a pattern matching method called as an associative pattern matching developed in our research on character recognition and attains a high accuracy in rough classification of words. The second part is concerned with a structural analyusis, that is, varification of the words based on phonem recognition.3. Adaptation to speakerWe have found the binary relations between Japanese vowels represented in their feature space that can be looked as almost independent on speakers and developed the vowel recognition technique mentioned above in 1.. An adaptation to speakers could be attained by using the vowel recognition technique in the process of verification of words given as candidates.4. Recognition ExperimentsThe word recognition technique was applied to 212 words produced by 10 mals speakers, and attained higher recognition rates that had everf obtained, that is, the percentage accuracy up to the 1st condidate was 98.2% and that up to the 2nd one 99.6%. However, these results was obtained by using the structural analysis basedc on the vowel recognition stated in 1. and the method of D.P. since our technique was not yet able to use phoneme recognition in the structural analysis. This is a feature problem.