Linear-Phase-Based Temporal-Filters-Optimization for Speech Recognition

[[abstract]]在人類文明中，語音辨識系統逐漸佔據我們的生活周遭，小至一般民眾攜帶3C的產品 ，大至專業領域的機台，都是語音辨識的各種運用，人們和機器之間的溝通必須藉由軟體轉譯而成，使軟體正確接收我們的意思然後傳達給機器，讓機器照著我們的指令做動作，這就是語音辨識系統的精神。 語音辨識系統正發展蓬勃中，目前許多專家.學者.廠商都致力於開創新的應用，而語音辨識的技術，套用在實際的真實環境時，辨識出來的效果經常與我們所預期的結果還不佳，主要是還有些技術需要克服。在這領域會遇到的瓶頸有:背景環境的噪音干擾、說話者語調的差異。 本篇論文主要著重於將語音類別分類做到完善，強健語音特徵參數；利用特徵參數在調變頻譜上的統計特性求取瞬間能量是種新的概念，引進中介向量將受限制的問題轉成不受限的問題，這裡採用線性鑑別分析法、最大類別距離法這兩種最佳化準則來求取濾波器係數，求取出來的濾波器係數將藉由HTK語音辨識系統來驗證其精確度。[[abstract]]In human civilization, speech recognition system gradually occupies our life. This phenomenon like as mall 3C products that everyone carried, specialty large-sized machine for researcher. These are speech recognition application. Communication of human with machine has dependent on translation software. We lead to software correctly receive our meaning, and then communicated to the machine. The machine followed our command. Speech recognition system in development is booming. Many specialist, master and businessman all endeavor to develop new application at the moment. But speech recognition technique used in natural physical environment always has low performance that over our expectancy. The main reason is some technology has to overcome. The bottleneck in this territory : Noise interference of background and difference of speaker intonation. This thesis focus on classify speech category and robust speech feature parameter, we used feature parameter statistics characteristic in modulation spectra to obtain instantaneous energy, it is a new concept. We used medium vector as a translation to turn the question limits. Two optimization norms : linear discriminant analysis and maximum class distance used to get filter output coefficient. The recognition accuracy experiment via HTK speech recognition system to tested.[[note]]碩