基于流形正则化极限学习机的语种识别系统

支持向量机 (Support vector machine, SVM) 在语种识别中已经起到了重要的作用.近些年来,极限学习机 (Extreme learning machine, ELM) 在很多领域取得了成功的应用.相比于 SVM, ELM 最大的优点在于极易实现、训练速度快,而且通常可以取得与 SVM 相近甚至优于 SVM 的识别性能. 鉴于 ELM 这些优异的特点,本文将 ELM 引入到语种识别中,并针对 ELM 由于随机初始化模型参 数所带来的潜在问题,提出了流形正则化极限学习机 (Manifold regularized extreme learning machine, MRELM) 算法.实验结果表明,在高斯超矢量(Gaussian supervector, GSV)特征空间上,相对于 SVM 基线系统,该算法对30秒语音的识别性能有明显的提升. 同时该算法也可以成功地应用到 i-vector 特征空间中,取得与当前主流的打分算法相近的识别性能.

Manifold Regularized Extreme Learning Machine for Language Recognition

Support vector machines (SVMs) have played an important role in the state-of-the-art language recognition systems. The recently developed extreme learning machine (ELM) which has been successfully applied to many areas tends to achieve much better generalization performance than the traditional SVM. Inspired by the excellent features of ELM, we introduce it into language recognition and propose a manifold regularized extreme learning machine (MRELM) to overcome the potential problem of ELM due to random initialization of model parameters. Experimental results show that the proposed algorithm can achieve much better performance than SVM at 30s durations in the Gaussian supervector (GSV) feature space. In addition, MRELM can be applied to the i-vector space and get comparable results to the existing scoring methods.