Utterance　verification　is　increasingly　essential　for　robustness　and　better　performance　of　speech　recognition　systems.　In　this　paper,　we　use　overlapping　one-class　SVMs　to　verify　utterances　and　propose　a　K-means　based　training　algorithm　for　the　overlapping　one-class　SVMs.　The　training　algorithm　first　divides　the　training　data　into　several　clusters　based　on　the　K-means　algorithm　and　then　expands　each　cluster　by　inserting　some　nearest　outside　data.　Then　it　iteratively　trains　the　overlapping　one-class　SVMs　on　the　expanded　clusters　and　constructs　the　train　clusters　based　on　the　learned　overlapping　one-class　SVMs　until　the　train　clusters　remain　unchanged.　Experimental　results　on　a　real　dataset　show　the　overlapping　one-class　SVMs　can　greatly　improve　the　recall　of　the　speech　recognition　systems.