Speech Feature Extraction Research Based On ISOMAP And Its Implementation On DSP

Currently, automatic speech recognition syste m (ASR) can obtain a highrecognition rate under laboratory conditio ns. But in the actual application environme nt,the recognition rate will s ignificantly decrease due to the influence of no ise. As themost critica l step in the procedure of the ASR, speech feature extractio n directlydecides the recognitio n performa nce of the syste m. With the widespread use ofportable devices, robust real-time ASRs in the mobile computing p latfor ms have hugemarket demands. This article investigates the e mbedded real-time ASRs, focus ing itsresearch on the real-time and robust speech feature extraction a lgor ithm.According to the nonlinear characteristics of speech s igna l and the characteristicthat Mel Frequenc y Cepstral Coeffic ients (MFCC) is unable to accurately represent thespeech signa l, this paper proposes the Samp le-ISOMAP based ma nifold hypothes is toimprove the MFCC algorithm. Furthermore, a real-time front-end processing speechsystem is imple me nted in the mob ile computing p latforms.MFCC has strong robustness and low algor ithm co mp lexity. However, theDiscrete Cosine Trans form (DCT) proc edure of MFCC is a pure ly mathe matica lca lculation based on the local optima l princ ip le and it has a high redundanc y. Thesethe DCT by S-ISOMAP transfor m, which mines the inherent nonlinear structura lfeature vector embedded in high-dimens io nal space and then reconstructslow-dimens iona l space to characterize speech signa ls.Then, using MATLAB and the Univers ity of Cambr idge deve loped speechrecognition toolk it (HTK), it build an HMM-based ASR as the experimental eva luationsystem. The evalution syste m is used to test and compare between the nove l and otheralgor ithms. The experimenta l results show that the proposed algor ithm has improvedthe robustness of ASR under different no ise environme nts and different SNRs.Fina lly, a high-perfor mance mobile computing platfor m—TMS320F28335DSPis selected and the front-end processing syste m based on the proposed algor ithm istransplanted to it, imp le menting a real-time fro nt-end speech processing syste m. In thesystem, the offline processing method is applied and the ISOMAP algorithm isimp le mented on a PC to train the temp late speech data. Then we get the S-ISOMAPtransfor mation matr ix stored in the DSP platfor m to extracting the S-ISOMAP-MFCCfeature parameters. The method ens ures the processing speed of the syste m and the rea l-time recognition.