A study of automatic transcription of music using a standard PC

Abstract

This thesis describes using a Personal Computer to identify notes that are played by a musical instrument. Several groups have been doing this work with more sophisticated laboratories and equipment with only moderate success. We have found the waves created by musical instruments vary, between instruments, a great deal in their stability and inherent vibration. It was more difficult to identify notes with very low frequencies than those with more central frequencies. We found it was very important to choose the correct starting point for the analysis with Fourier Transform otherwise we would not be analysing the stable stage of the wave. We tried simple strategies to initially reduce the number of computer operations, and memory requirements, with marginal success. We followed with more complex subtraction strategies which were much more successful. The most useful technique involved creating a "calculated percentage multiple" which was almost 100% successful in identifying single notes. For multiple notes we were surprised to find that a group of five different instruments from MIDI were a better source of "known" notes to compare with the "unknown" notes than the MIDI equivalent of the real instrument playing the music. These methods were developed using midi instruments but were verified using a real grand piano. We suggest some further lines of enquiry that may make this technique more successful.