计算机系统应用
計算機繫統應用
계산궤계통응용
APPLICATIONS OF THE COMPUTER SYSTEMS
2014年
9期
6-12
,共7页
声纹识别%矢量量化%优势节点树%高斯混合模型%通用背景模型
聲紋識彆%矢量量化%優勢節點樹%高斯混閤模型%通用揹景模型
성문식별%시량양화%우세절점수%고사혼합모형%통용배경모형
voiceprint recognition%vector quantization (VQ)%vantage point tree (VPT)%gaussian mixture model (GMM)%universal background model (UBM)
目前声纹识别系统已经实现较高的识别精度,但是随着目标说话人个数的增加,一般系统很难满足实时性的要求,由此提出一种双层识别模型。在第一层识别模型中,采用基于VQ-VPT(Vector Quantization-Vantage Point Tree)模型进行快速匹配,挑选出与测试者声纹特征最相近的K个目标说话人声纹模型。在第二层识别模型中,采用GMM-UBM(Gaussian Mixture Model-Universal Background Model)模型,精确匹配上层模型得到的K个目标说话人声纹模型,并做出最终的判决。实验验证,双层识别模型在确保高识别精度的前提下,大幅度的提高了系统的识别速度。
目前聲紋識彆繫統已經實現較高的識彆精度,但是隨著目標說話人箇數的增加,一般繫統很難滿足實時性的要求,由此提齣一種雙層識彆模型。在第一層識彆模型中,採用基于VQ-VPT(Vector Quantization-Vantage Point Tree)模型進行快速匹配,挑選齣與測試者聲紋特徵最相近的K箇目標說話人聲紋模型。在第二層識彆模型中,採用GMM-UBM(Gaussian Mixture Model-Universal Background Model)模型,精確匹配上層模型得到的K箇目標說話人聲紋模型,併做齣最終的判決。實驗驗證,雙層識彆模型在確保高識彆精度的前提下,大幅度的提高瞭繫統的識彆速度。
목전성문식별계통이경실현교고적식별정도,단시수착목표설화인개수적증가,일반계통흔난만족실시성적요구,유차제출일충쌍층식별모형。재제일층식별모형중,채용기우VQ-VPT(Vector Quantization-Vantage Point Tree)모형진행쾌속필배,도선출여측시자성문특정최상근적K개목표설화인성문모형。재제이층식별모형중,채용GMM-UBM(Gaussian Mixture Model-Universal Background Model)모형,정학필배상층모형득도적K개목표설화인성문모형,병주출최종적판결。실험험증,쌍층식별모형재학보고식별정도적전제하,대폭도적제고료계통적식별속도。
At present, the voiceprint recognition system has achieved high identification precision. But with the increase of the number of target speakers, general system has difficulty in satisfying the need of real time. Therefore, a two-layer recognition model is raised in this paper. The first layer based on VQ -VPT model quickly sorts out K target speakers’ voiceprint models which are most similar to the speaker’s voiceprint characteristics. In the second layer, the GMM-UBM model matches the K voiceprint models to make a final judgment. Via experimental verification, under the premise of ensuring high recognition accuracy, the two-layer recognition model has greatly improved the recognition speed of the system.