A research of improved algorithm for GMM voiceprint recognition model
Jing Zhang, Xiaomei Chen · 2016
As an economical, reliable, easy and safe way to recognize identity, the Voiceprint recognition has a bright prospect in the market at present. However, the bad robustness in a noisy environment constraints its practical application, so the key make voiceprint recognition be applied widely is to improve the robustness of system. A Voiceprint recognition system is mainly made up of signal recognition,feature extraction, model building and matching, currently, the study to improve the recognition rate is mainly based the technology just mentioned. The paper studied a method to improve the system robustness by improving the voiceprint model. The GMM is text-independent model and used widely. In order to improve the system recognition rate, this paper improved GMM model during training and recognition period respectively. During training period, a k-means algorithm based on adjacent rules was put forward to obtain initial GM M value, which overcame the flaw that the system employs poor overall performance due to excessive attention on several indicators by the conventional method. It increased training speed and recognition rate by virtue of simplifying the derivative process of EM. During recognition period, in order to avoid the influence of bad frames on the final result, a frame matching weight method based on entropy was proposed. The experiment result shown that improved GMM model effectively increased the robustness of voiceprint recognition system.