The　accuracy　of　speech　recognition　directly　depends　on　accurate　endpoint　detection.　Endpoint　detection　is　a　very　difficult　task　in　the　noise　environment.　It　will　be　degraded　with　the　decrease　of　SNR　and　different　noise　affects　the　accuracy　of　speech　recognition.　As　a　result,　this　paper　proposed　an　endpoint　detection　approach　which　is　applicable　to　the　telephone　speech　recognition　system　for　city＇s　name.　The　approach　integrates　band-partitioning　spectral　entropy　and　spectral　energy　to　form　a　set　of　new　feature　parameters　that　can　compensating　the　drawback　of　entropy　and　energy　so　that　the　performance　of　the　detection　is　improved.