【摘要】 针对大间隔高斯混合模型基于LBFGS参数更新算法收敛速度慢的不足,提出一种快速参数更新算法。采用构造弱意义辅助函数的方法,得到扩展Baum-Welch算法形式的快速参数更新公式。利用大词汇汉语语音库上的声调分类任务来验证训练速度与分类性能。实验结果表明快速参数更新算法只需数次迭代就能收敛至最优结果,较LBFGS优化方法在识别性能相当的情况下具有更快的训练速度。更多还原

【Abstract】 To speed up training of large margin Gaussian mixture model based on LBFGS optimization routing, a rapid model parameter method is proposed. The method formulates extended Baum-Welch algorithm like updating equations by constructing weak-sense auxiliary function. The proposed algorithm is experimented on Mandarin tone recognition tasks. It is shown training can be accomplished within only several iterations, much faster than that of LBFGS optimization routine.更多还原