We describe a method of lexicon expansion to tackle variations of spontaneous speech. The variations of utterances are found widely in the programs such as conversations talk shows and are typically observed as unintelligible utterances with a high speech-rate. Unlike read speech in news programs, these variations often severely degrade automatic speech recognition (ASR) performance. Then, these variations are considered as new versions of original entries in the ASR lexicon. The new entries are generated based on the SMT approach, in which translation models are trained from corpus translating phoneme sequence in a lexicon into the sequence obtained by phoneme recognition. We introduce a new method in which unreliable entries are removed from the lexicon. Our SMT-based approach achieved a 0.1 % WER reduction for a variety of broadcasting programs.