Sound Source Separation in Application to Singing Voice Extraction from 2-Channel Polyphonic Musical Recordings

Sound source separation is of great interest in the field of ”musical information retrieval“ and as such in focus of many recent studies. Generally speaking, the goal is to identify and/or extract the different musical sources of a recording, e.g., instruments, vocals etc. Having separated sources, further processing can be made, e.g., transcription. This work concentrates on vocal extraction in order to provide a basis for advanced tasks, e.g. lyrics recognition or singer identification, which will not be part of this study. Once the singing voice is separated it will be removed from the original signal to yield the accompaniment signal which itself can be used for further information retrieval purposes or for Karaoke applications. The design and implementation of a suitable method to separate singing voice and accompaniment, which may include probabilistic models (Hidden Markov Models HMM, Gaussian Mixture Models GMM etc.) and/or transcribing the vocal melody, is examined.