This paper discusses musical form from a cognitive and a computational viewpoint. While several time-windows exist in the brain, we here put emphasis on the superchunks of up to more than 30 seconds lengths. We compare a strategy for auditive analysis based on human cognition with a strategy for automatic analysis based on feature extraction. The feature extraction is based on the musical features rhythm, timbre and chroma. We then consider the possible consequences of this approach for the development of music generating software.