基于WFST的字素因素转换（g2p）：开源组件

概要：support EM sequence alignment（最大期望序列），与RNNLM结合perform well。同时也展示了Lattice Minimum 贝叶斯译码(LMBR)。Toolkit用Openkit运行。

简介：（Spanish and Italian) rule of them are consistent, it is straightforward for conversion. however, it is more complex for English and French. 

structure:

scetion 2:background

g2p probelm including:

1,sequence alignment(align the sequence between grapheme and phoneme)

2,model train

3,decode

section 3:alignment approach

WFST(feature:EM-drived,mutiple to mutiple)

modifications including:

1,while training: considering m2one nad one2m.

2,construct lattice and delete unconnected arcs.

3,non zero.

section 4:joint-sequence LM(language model)

procedure

1,

2,

3,N-gram ->WFST(applied in all language)

section 5:decode approach

音素格子的最短路径

section 6:previous experiment

文中提出的m2m-fst-P和rnnlm-P能够达到高水平。

section 7:usage and conmmand

NETTalk的例子

term:data-drived(数据驱动),state of the art(达到高水准的),lattice(格子框架)，ASR（automatic speech recognization),

文中用到的测试字典：NETTalk、CMUdict和OLAD（牛津进阶）