一种基于SDTS的HMM训练算法

A HMM Training Algorithm Based on SDTS

用传统的BW算法训练语音识别系统的HMM需要大量的语音数据。本文在假设声学模型系统的子空间捆绑结构(SDTS)为己知的前提下,提出了一种新的训练算法,可以有效地减少系统对训练数据的需求。理论分析和仿真表明,与传统的BW算法比较,新的训练算法(IBW)可压缩模型参数15倍,从而可大量地减少训练数据。尽管新算法要用到系统的先验知识,但它还是显示了许多优越性。

It generally requires a large number of speech data for a speech recognition system to train HMM by the BW algorithm. In tin's paper we devise a new training algorithm (we call it IBW algorithm) with the aim to reduce some of training dara. assuming on a prior knowledge of the subspace distribution tying structure of the system (SDTS). Our computer simulation and theoretical analysis show that IBW algorithm can reduce model parameter roughly 15 times and decrease requirement of speech data to train HMMs but with no loss in word accuracy compared with traditional BW algorithm.