The default Phonetisaurus approach can probably be described as a mixture of the EM approaches in Jiampojamarn's m2m-aligner, and the Kubo aligner (although it precedes the latter). Kubo introduced a better solution for managing the multigram tokens, particularly for languages like Japanese. This involves penalizing the arc weights during training. It should be possible to implement this in the new aligner using an alternative ArcMapper, and add this as an option.
The default Phonetisaurus approach can probably be described as a mixture of the EM approaches in Jiampojamarn's m2m-aligner, and the Kubo aligner (although it precedes the latter). Kubo introduced a better solution for managing the multigram tokens, particularly for languages like Japanese. This involves penalizing the arc weights during training. It should be possible to implement this in the new aligner using an alternative ArcMapper, and add this as an option.