語音辨識系統與單聲道消噪之DSP實現

Abstract

在近年來的科技進步之下，語音辨識的功能已經是許多3C產品的必備介面；現今亦可看見許多語音辨識的應用，例如車用免持手機系統，智慧型手機，GPS等產品，本論文在語音辨識的部分是取出梅爾特徵參數之後再採用動態時軸校正的方法比對並且辨識，並且提出使用單聲道噪音消除的方式，來改善有噪音的語音訊號，使語音訊號能有更好的音質，最後將單聲道噪音消除應用在提升語音辨識系統的辨識率上，並且測試單聲道噪音消除對語音辨識所造成之影響。 本論文是使用數位信號處理器TMS320C6416實現語音辨識系統和單聲道噪音消除系統，在考量數位信號處理器的運算能力和語音參數的資料型態(數位信號處理器可能未搭載浮點運算器)的情況之下，我們使用整數型態替代浮點型態來表示語音相關參數，並且調整語音相關參數，使此系統運算能達到最佳即時(Real-Time)的效果。

With the scientific and technological progress in recent years, the speech recognition interface has been essential in many 3C products, such as hands-free car kit, smart phones, GPS, and other products. This paper will propose a speech recognition system, consists of speech enhancement algorithm, speech feature extraction, and recognition method, which employs recognizes the words of command voice. For the purpose of increasing the purity of the command voice, we adopt single-channel noise reduction (NR) as our speech enhancement algorithm. The speech feature extraction applies mel-scale frequency cepstral coefficient (MFCC). We employ dynamic time warping (DTW) to do dynamic matching after feature extraction. In experiment, we use the DSP TMS320C6416 to implement this speech recognition system. We investigate the computation of floating-point MFCCs to fixed-point ones. In particular, we have to check the scaling factors during each stage of the computation of MFCC. These scaling factors are carefully chosen such that the highest precision is achieved without overflow. The results of experiment show that the recognition rate will be improved at low SNR case.