In　English　phonetic　synthesis,　it　is　impossible　to　create　a　thesaurus　containing　all　vocabulary　as　English　has　an　almost　unlimited　vocabulary.　Hence,　for　English　words　that　are　not　included　in　the　thesaurus,　generating　phonetic　symbols　through　the　Grapheme-to-phoneme　(G2P)　algorithm　is　the　best　solution.　For　this　purpose,　a　dynamic　finite　generalization　(DFGA)　machine　learning　algorithm　for　the　rules　of　G2P　conversion　is　proposed　in　this　paper.　The　dictionary　library　used　for　learning　has　27,040　words,　90%　of　which　are　used　for　rule　learning,　and　the　remaining　10%　are　used　for　testing.　After　ten　rounds　of　cross-validation,　the　average　grapheme　conversion　accuracy　in　the　learning　and　test　sets　is　99.78%　and　93.14%,　and　the　average　vocabulary　conversion　accuracy　is　99.56%　and　73.51%,　respectively.　©　2020　Published　under　licence　by　IOP　Publishing　Ltd.