| In this paper, voiceprint recognition system consists of three parts which areendpoint detection, feature extraction and model matching. It is studied doublethreshold method based on linear prediction cepstrum distance and short-timezero crossing ratio in endpoint detection part.The traditional double thresholdmethod can’t detect low energy speech segment, but the new double thresholdmethod can solve the problem through experiments. In the feature extraction part,the combination of mel frequency cepstrum coefficient and differential frequencycepstrum coefficient reflects the speaker’s personality better. Then it’s proposed anew model parameters initialization method combining division algorithm withk-means clustering algorithm after studying the gaussian mixture model. Someexperiments about two kinds of endpoint detection algorithm, feature extractionand training method using gaussian mixture model have been made. The experi-ments show the whole recognition system having high recognition rate under theenvironment of pure voice.One of the difficulties in voiceprint recognition research is the research ofrecognition system under the background of noise. Although the system gets pe-ferct performance under the environment of clean speech recognition system, therecognition rate is decreased obviously in the noise environment. In this paper,denoising processing experiments using the wavelet transform and waveletpacket transform show that the effect of wavelet packet denoising is obviouslybetter than the wavelet transform. Then it’s proposed a improved threshold de-noising method based on the common wavelet packet threshold and compromisethreshold. The new algorithm of this paper removes noise well through compar-ing the simulation experiments of speech signals and the experimental data. Andthe denoising recognition system has achieved higher recognition rate. At last,this algorithm is applied to speech signal with the actual complex noise, and getrid of the noise effectively. |
PDF Full Text Request