A Study of Distinctive Feature Theory for Perception and Cognition and its Application to the Engineering Field

在本课题的最后一年，我们在特征理论及其应用方面取得了实质性的进展。声学特征的建立：我们从声学语音解码的内部层次结构上，详细研究了声学语音解码的独特特征（这里我们称之为声学特征）。回到布卢姆菲尔德学派的不变性理论，我们研究了基于声学特征的语音识别范式的语音和音系可行性。听觉机制和声学特征系统的底层结构：人类感知特征与机械识别特征（即贝叶斯分类器）的对比。一个合理的假设是，我们的声学特征系统支配着听觉系统。基于听觉模型的声音过滤：我们改进了耳蜗和毛细胞的模型，与许多生理数据非常吻合。该模型在爆炸点提取和VOT测量方面优于传统方法。我们利用多个dsp开发了该模型的实时处理器。用于人工耳蜗的声学处理器：为了听到语音处理，我们将声学模拟器直接耦合到可植入的接收器-刺激器单元。然后在模拟器上附加一个与当前模拟处理器完全兼容的新数字化处理器板。健康的人可以理解以前从未经历过的严重扭曲的语言，而独特的特征可以分解误解。通过与先前关于植入患者表现的报告的合理一致，我们确认了进一步发展新的语音处理策略的基本基础。其他语言的应用：我们的声学特征的有效性不仅限于日语，而是推广到其他语言。同样的框架成功地适用于法语。

This final project year, we have progressed substantially in the distinctive feature theory and its applications.1. Establishment of acoustic features : We studied the distinctive features applicable to the acoustic phonetic decoding (here we call this the acoustic features) in detail on its internal hierarchical structures. Back to the invariance theory of the Bloomfield's School, we examined both the phonetical and phonological feasibilities of the acoustic feature based speech recognition paradigm.2. Underlying structures across the auditory mechanism and the acoustic feature system: Human perceptual characteristics contrast with mechanical recognition characteristics, i. e. the Bayesian classifier. A reasonable hypothesis was that our acoustic feature system rules the auditory system.3. Auditory model based acoustic filtering : A model for the cochlea and the hair-cells, we have refined, agreed remarkably with a lot of physiological data. The model outperformed over conventional methods in extraction of the burst point and measurement of VOT. We developed a real time processor of this model using multiple DSPs.4. Acoustic processor for the cochlear implant : In order to hear speech processing, we coupled an acoustic simulator directly to the implantable receiver-stimulator unit. Then to the simulator attached a newly digitized processor board fully compatible with the current analog processor. Healthy people could understand heavily distorted speech never experienced before, and the distinctive features could decomposed the misperceptions. Qualifying the reasonable agreements with previous reports on performances of implanted patients, we have confirmed the fundamental basis for further developmeny of new speech processing strategies.5. Application to other languages : Effectiveness of our acoustic features is not limited to Japanese, but is generalized to other languages. The same framework was successfully applicable to French .