The Research Of High-level Information Fusion Based Speaker Recognition Algorithm Using Short Utterance

Speech is an important carrier of human affective interaction and shared cognition,and is also the most basic and natural communicative means.Speaker recognition,also known as voiceprint recognition,is a subject that studies how to extract personal information from speech waveform,and then take use of the individual information to complete recognition task through modeling.Gaussian Mixed Model is widely applied in the field of speaker recognition for its excellent performance.In the GMM-UBM system we use a single,speaker-independent background model to represent the impostors in order to solve the mismatch problem between test environment and training environment.In the UBM-MAP-GMM system,we derive the hypothesized speaker model by adapting the parameters of the UBM using the speaker's training speech and a form of Bayesian adaptation.In the meantime,scholars pay more attention to prosodic features due to the extraction of short characteristic parameters gradually entered into bottleneck.In addition,compensation technologies are put forward constantly to solve the channel mismatch problem.JFA and i-vector with pragmatic theories are more popular.In response to above analysises,the main research and results are relected in the following aspects:1.In order to solve the problem that parts of the Gaussian components have low contribution to score calculation,we propose a sorted Gaussian component based speaker verification method.It is generally considered that the components in GMM represent the phoneme classes,while the components in UBM represent the whole phoneme classes.We derive the speaker-dependent GMM by adapting the parameters of the UBM using the speaker's training speech and a form of Bayesian adaptation.In fact,the adapted GMM include both personal information and non-target speaker information.It is also worth considering that the tranning data are limited and parts of phoneme class information are insufficient,wich lead to low discrimination in score calculating.The experimental results show that the improved system has good performance in time complexity,equal error rate,and the physical storage space.2.The short-term characteristic parameters encountered a bottleneck in the aspect of improving performance.Based on this background,we put forward high level information fusion based speaker verification algorithm.This adapted algorithm fused short-term characteristic parameters with high level prosodic features together under secondary judgment mechanism.The short-term characteristic parameters reflect the vocal tract information while the prosodic features reflect the glottis information.So they can improve the system performance in theory.The experimental results show that the high-level information fusion based text-dependent speaker verification system can reduce the equal error rate effectively.3.In this part,we extracted mean Gaussian supper vector form special adataed personal GMM.We have found in trials that the general information is included in the front of few dimensions of GSV after feature dimension reduction processing.In the meantime,we found that the personal information is included in the specific dimensions after PCA process.Based on the former,we proposed gender discrimination strategy and GMM_SUBM framework.Based on the latter,we put forward a PCA based spearker verification algorithm.Under this framework,we extracted the identity vector from the utterances and used cosine score.Experimental results show that PCA based spearker verification algorithm can reduce the system equal error rate and improve system performance.