Musical Blind Source Separation Using Ambisonics

This thesis deals with the musical blind source separation problem: given multiple instrument tracks recorded together, how can each be isolated from the others given no additional knowledge about the instrument locations or sounds? There are many advantages to solving the problem: live performances can convey the energy of the performers, and the sound of the room can improve the perceptual quality of the recording, while separating the instrument tracks afterwards allows for more pre-cise equalization and mixing to improve the recording. This thesis proposes a novel approach to the problem using an ambisonic microphone array. The directionality of the microphones in the array provides information about the location of each source and allows for the implementation of a direction of arrival estimator that calculates the position of each source based on the directions that receive the most non-reverberant acoustic energy. Given the directions of arrival (DOAs), the method performs spatial filtering to virtually steer the microphone array in the desired direc-tions. This approach is compared with a classic DOA estimator, MUSIC, a classic non-negative matrix factorization (NMF) approach, and a state of the art ambisonic domain filtering approach. The proposed method outperforms NMF and MUSIC in nearly every tested configuration, and can outperform the ambisonic filtering ap-proach in certain high-reverberation cases. The proposed method is significantly more computationally efficient than the comparison methods, working much faster while introducing less algorithmic noise to the separated tracks. The results raise the question of how to best balance performance and computational complexity in different use cases.