透過您的圖書館登入
IP:18.221.27.155
  • 期刊

建構在ADSP-2181上的語音辨識器

Realization of Speech Recognition on The ADSP-2181

摘要


在本論文中的語音辨識器是非特定語音辨識系統且具抗噪能力,我們會介紹如何強化MFCCs成為強健性特徵參數。語音辨識是一個複雜的技術,需要許多複雜的數學運算。為了讓語音辨識技術可以在定點的DSP上執行,我們除了要定點化快速傅立葉轉換、計算自相關係數、Durbin's algorithm等萃取特徵參數時使用的數學運算,還要考慮語音模型的個數、語音特徵參數的維度、辨識率及辨識時間等因素。我們採用Q-format來做定點化,並根據辨識速度、辨識率與記憶體使用空間來決定語音模型的數量和語音特徵參數的維度。非特定語者且具有抗噪能力的語音辨識技術,可用於聲控玩具、手機上的聲控撥號與汽車上的語音命令控制。

並列摘要


The speech recognizer that we develop is speaker independent and robust in noise environment. We explained how to develop robust speech feature and how to realize the speech recognizer on ADSP- 2181. In order to realize speech recognition technique on fixed-point DSP, we must use fixed-point arithmetic instead of floating-point arithmetic and consider hardware resources. We use Q-format technique to complete fixed-point arithmetic. Considering execution time, recognition accuracy rate, and memory size, we investigate how to reduce the amount of acoustic model and dimension of speech features. The robust speech recognition technique can be used in some applications, such as voice dialing and voice command controlling in car.

並列關鍵字

ASR Fixed-point DSP

延伸閱讀