Abstract: The primary task of audio clip recognition and retrieval is to construct a well performance classifier learning machine How to choose informative training instance from redundant training database and reduce training time of classifier machine is a challenge during the construction of classifier machine, especially for audio clip recognition with large size training database Since support vector is the key instance in support vector machine （SVM）, an algorithm to train SVM with incremental learning is proposed In this algorithm, training database is segmented into sub databases and each sub database is trained in batch During each training process, only support vector is reserved for future training and non support vector is discarded Compared with traditional and decremental SVMs, this training algorithm obviously reduces training time and obtains high correct rates of recognition and retrieval