Implementation and Improvement of Integer-type FFT for Speech Recognition

• Main Authors: Yu-Chang Chou, 周俞璋
• Other Authors: Jyh-Shing Roger Jang
• Format: Others
• Language: zh-TW
• Published: 2007
• Online Access: http://ndltd.ncl.edu.tw/handle/82380555794124268939

碩士 === 國立清華大學 === 產業研發碩士積體電路設計專班 === 95 === In this thesis, we investigate the methodologies of porting the computation of floating-point MFCCs to integer ones. In particular, we focus on the implementation of integer-FFT during the integer-MFCCs computation. We have closely checked the scaling-up factors during each stage of the computation of integer-FFT by using a data-driven approach. These scaling-up factors are carefully chosen for striking a balance between precision and overflow, such that the highest recognition rate can be achieved. Moreover, we have proposed the use of some characteristics in trigonometric functions in building a minimum lookup table without degrading the recognition rate.