'With the increase of music objects available, there is a need for searching similar songs. This thesis proposes and investigates two novel approaches for audio similarity. One is a non-feature based algorithm called Normalized Compression Distance (NCD), which uses a distance to tell audio similarity using compression technique. The other is amplitude based algorithm, using Amplitude Pattern Mining (APM) to measure similarity. It is found that NCD is not applicable because of the incapability of real world compressors, amplitude is a promising feature and APM shows good performance. For both distance and APM approaches, models are built for playlist generation. '