Audio　classification　is　the　basis　of　content-based　audio　analysis　and　retrieval.　The　conventional　classification　methods　mainly　depend　on　feature　extraction　of　audio　clip,　which　certainly　increase　the　time　requirement　for　classification.　An　approach　for　classifying　the　narrow-band　audio　stream　based　on　feature　extraction　of　audio　frame-level　is　presented　in　this　paper.　The　audio　signals　are　divided　into　speech,　instrumental　music,　song　with　accompaniment　and　noise　using　the　Gaussian　mixture　model　(GMM).　In　order　to　satisfy　the　demand　of　actual　environment　changing,　a　universal　noise　background　model　(UNBM)　for　white　noise,　street　noise,　factory　noise　and　car　interior　noise　is　built.　In　addition,　three　feature　schemes　are　considered　to　optimize　feature　selection.　The　experimental　results　show　that　the　proposed　algorithm　achieves　a　high　accuracy　for　audio　classification,　especially　under　each　noise　background　we　used　and　keep　the　classification　time　less　than　one　second.