The aim for this project was to research, design, and provide the basis of an implementation for a real-time melody recognition system. The system should match music hummed or sung by a human with a particular piece in a melody database. The system should also allow for very large melody databases, where linear searching of the database would be too slow. The design of the system consists of the following components: • Digital sampling, • Note recognition, • Melody recognition, and • Melody indexing. They are linked together in a serial manner, the output of one component being the input to the next. The overall structure of the system is shown in Figure 1.1. Some of the system components have been well researched (such as note recognition [Kuh90],[Lan90),[SJ89],[Ric90]) and others have not (such as melody recognition). The main thrust of this report relates to melody recognition and how this effects the indexing of the melody database. An important part of this is experimentation with various possible melody correlation functions. A good function should have the following three characteristics

• accuracy (melodies that almost match should be given much higher correlation values than those that do not), e speed (essential if the system is to exhibit real-time performance), and • index-ability ( easy to modify to allow indexing, so matching can be done in less-than-linear time with respect to the database size). Most project experiments were performed on a Sun SPARCstation 10, using a melody recognition test program called mrs, which is written in C.