The　Mel-frequency　cepstral　coefficients　(MFCCs)　based　on　human　auditory　characteristics　are　widely　used　for　audio　recognition.　However,　the　performance　of　MFCC-based　audio　recognition　degrades　due　to　noise　interference.　In　consideration　of　this,　we　propose　the　matching　pursuit　(MP)　sparse　representation　algorithm　based　on　genetic　algorithm　(GA)　improved　by　elite　strategy　and　evolution　reversal　to　accomplish　the　task　of　filtering　out　extraneous　noise.　In　the　first　step,　MP　is　carried　out　to　represent　the　ecological　environmental　signal＇s　inner　structure.　The　second　step　consists　of　MFCCs　feature　extraction.　Finally,　two　different　classifiers,　Support　Vector　Machine　(SVM)　and　Gaussian　mixture　model　(GMM)　were　performed　and　compared　using　the　proposed　features.　Experimental　results　showed　that　the　SVM-based　classifier　outperforms　the　GMM　classifier　and　indicated　that　this　method　with　sparse　representation　achieved　improved　performance　in　noisy　environments.　©　2012　IEEE.