Research On Speaker Recognition Of The Robustness Based On I-vector

Speaker recognition is the ability to allow the machine to distinguish the different sounds of people, mainly in the voice signal as the research object, which feature extraction and pattern recognition by the two major components; wherein, how to effectively extract the personality traits distinguish different speakers have been since the study is the difficulty.In the absence of noise conditions speaker recognition system recognition rate is quite high, but under noisy conditions its stability will be a sharp decline, how to improve the noisy speech speaker recognition is particularly important.This article is based on Kaldi speech recognition tool as a platform for simulation experiments, an open source tool written in Kaldi by Cambridge University with a C ++ package, easy to modify and expand depending on the purpose of the experiment. Paper first introduces the basic knowledge of speaker recognition system, and then describes in detail the current mainstream speaker recognition technology i-vector, combined with voice recognition tools Kaldi speaker recognition system training parameters were optimized.Speaker recognition rate for the case of noisy conditions, we use noise from the encoder and the i-vector combination of methods to improve the recognition rate. We know that between noise and speech signal has a very complicated relationship, but in real life mainly additive noise, hence we only studied the Gaussian white noise on system stability, it has set 0dB, 5dB, 10 dB, 15 dB, 20 d B, with the noise from the encoder signal noise suppression, to retain the speaker’s characteristic differences, thereby increasing i-vector speaker recognition system robustness. Experimental results show that the denoising i-vector recognition rate significantly improve.