@article{oai:bunkyo.repo.nii.ac.jp:00003258,
 author = {鈴木, 昇一},
 journal = {情報研究, Information and Communication Studies},
 month = {2002-01-01, 2011-02-23},
 note = {音声・音楽の処理技術を確保するのに,隠れマルコフモデルの理論(確率的有限オートマトンの理論)が基本的に重要な役割を果たすようになって,久しい.こうも隠れマルコフモデルの理論の全盛時代が長期にわたって続くと,これを上回り打ち破る技術が登場することが望まれることになる.本論文では,隠れマルコフモデル[A5]を適用して得られる在来の技術に対抗するため,文脈(音素近傍)を考慮し,音素を認識する技術をSS理論[B3],[B4]を適用して,構築する.
　処理の対象とする問題のパターンを多段階にわたって変換しながら認識する手法は魅力的である.多段階認識法の1例として,SS多段階想起認識法がある.SS多段階想起認識法は,多段階にわたってパターンモデル変換を行い,構造受精変換の不動点(あるカテゴリの代表パターンのモデル)を探索する形で想起し,認識する手法,つまり,不動点探索型・多段階パターンモデル帰納推理変換・想起認識法である.具体的には,本論文では,SS多段階想起認識法で使用できるモデル構成作用素T,類似度関数SM,大分類関数BSCを文脈(音素近傍)を考慮し,構成する.パターンの帰属する有効な候補を絞り込む機能を持つカテゴリ選択関数CSFはSM,BSCを用いて構成できる. T,SM,BSC,CSFは各々,axiom 1～4を満たすものである.
　また,Parzen Window法,楕円体特徴間距離,重み付きDice係数,Jaccard係数,2次ニューラルネット,階層型2層ニューラルネットを用いて,SM,BSCが構成され得ることも示される.
　画素近傍を考慮し,画像を認識するシステムを構築するためのT,SM,BSCをも提案している.
\n　A theory of a hidden markov model(a theory of a probabistic finite automaton) has been played for a long time a fundamentally important part in the synthesis and recognition of speech and melody. It is desirable that the theory should be removed as soon as possible because we have its best long days until now. A technical method of phones against old-fashioned techniques obtained with help hiddon Markov models [A5] is presented here paying attention to parts directly before or after a phoneme in question and which influence its meaning, which is built up by making use of S.Syzuki theory.
　Procedures of successively transforming an input pattern to be recognized in question into a prototypical pattern offer an appeal to us. There is SS-method of multistage associative recognition as an example of such multistage recognition methods.
　SS-method is dissolved into five steps:
　(1) To call for a corresponding model of an input pattern.
　(2) To one after another transform it into pattern-models inductively.
　(3) To seek for a fixed-point of a structural fertilization change.
　(4) To determine a category to which the fixed-point belongs.
　(5) To determine a category of the input pattern to be its category.
　That is to say, SS-method is explained as associatively recognition of an input pattern by transforming inductively pattern-models throughout multistage and searching out a fixed-point.
　We shall embody a suitable arrangement of model-construction operators Ts, similarity-measure functions SMs and rough classifiers BSCs which can extract the meaning of an input pattern from its context for use of SS-method. If SM and BSC is given, we can design a category-selection function CSF which extract a list of significant categoies to one of which a pattern in question may belong. T, SM BSC, CSF must satisfy axiom 1～4 respectively.
　Moreover we construct SM and BSC by making use of a method of Parzen window, an ellipsoid feature-distance, a weighted Dice coefficient, a Jaccard coefficient, a second-order neural network, and two-layer feed-forward neural network.
　T, SM and BSC are proposed here for a system which recognizes an image considering a neighbour of each pixel.},
 pages = {69--141},
 title = {近傍を利用した音素認識のためのモデル構成作用素T,類似度関数SM,大分類関数BSCの諸構成と,SS不動点探索型多段階想起認識},
 volume = {28},
 year = {}
}