國語單音節即時辨認系統之製作

Abstract

本論文之製作--即時國語單音節辨認系統，是以遞迴式類神經網路模型為 辨認的方法，對411音做單字音辨認系統。如包含四聲聲調的辨認，則 為1280音。由於個人電腦的運算速度仍不夠快，因此由個人電腦搭配運算 速度較快的DSP board來完成，辨認的工作用 DSP board 完成計算，並將 結果傳到個人電腦顯示。對於即時辨認系統在製作上，我們希望它辨認速 度愈快愈好，因此能否充分的運用週邊源，非常重要。在流程安排上，即 時系統有許多工作必須及時處理，因此工作流程上我們安排取樣、參數抽 取、邊界判定、語音辨認等工作平行處理，使所有工作都能及時處理。另 外在前處理方面，本實驗以二種不同方式為前處理應用：一為單字音邊界 判定法，另一為有限狀態機（Finite state machine）。前者的好處是可 以準確的抓出字的有效範圍，避免包入太多的低能量範圍區域進入辨認； 後者的好處是不必往回尋找有效範圍，對每個音框(frame)直接做取捨的 判定，並經由此狀態的判斷，為日後的連續音辨認前處理做準備。 The design and real-time implementation of a recurrent neural Network (RNN) based speech recognizer for 1280 isolated Mandarin syllables is presented in this thesis. The recognizer adopted the RNN-based recognition method because of its good performance evaluated based on off-line simulations. The system is realized on a TMS320C40 DSP module added on a PC/AT computer. Two versions using different pre-processing schemes have been implemented. One first uses an endpoint detector to extract the speech part of the input signal. The recognition process then starts at the moment when the ending point is detected. The processing delay is significant because of the sequential processings of the input signal. The other first uses a finite state machine to categorize the input signal, frame-by-frame, into 8 states including silence, vowel, nasal, fricative, and their following transient states. State- dependent constraints are then added to restrict the search of optimal path in order to speed up the modified one-stage recognition process. Besides, the recognition process can be run in parallel with the finite state machine. So it runs much faster than the first version. No perceptible processing delay can be found. A recognition rate of 76.2% was achieved.