Commercial programs exist that can successfully detect the pitch of an instrument vs. time, but recordings of human vocals present unique issues due to their noisy frequency spectrum. The goal of this research is to design an algorithm capable of determining pitch vs. time for a single-track, monotonic, audio recording of human vocals. The algorithm is based on a combination of Windowed Fourier Transform and Artificial Neural Network processing. The pitch vs. time information is stored in a MIDI file that is effectively a reproduction of the original vocal recording. The thesis discusses the algorithm and its implementation in the MatlabRTM environment. Examples of applications and experimental results are also proposed.
展开▼
