By analyzing time-variant frequency, magnitude and phase of harmonics derived from the musical timbres of acoustic and electronic instruments and synthesizers, obtain a set of features that are adequate for timbre description and sound synthesis. Principle controlling parameters of timbre can be later determined. Map the semantic descriptions that people use on timbre to the scales of these features, and with ML algorithms, characteristics of musical instruments are also attainable. Finally, a new approach of timbal sound design by assigning the parameters and forming the spectrum will be developed. [Details of the Thesis] [Video]
Computer-assisted auto-orchestration and texture generation
This project aims at the generation of symbolic representation of multi-track orchestra music given the constraints of orchestra configuration, desired morphology of the sound, and the measure of textural complexity. By studying from the dataset of paired music chunks and scores of orchestra music, the neural network learns the latent orchstration knowledge and generate musically reasonable scores that to the best meet the conditions.