|
|
Application for Guitar Sound Separation from Music Recording
Holková, Natália ; Rohdin, Johan Andréas (oponent) ; Mošner, Ladislav (vedoucí práce)
This thesis aims to implement a model capable of separating guitar sounds from a recording and use it in a practical application. It was necessary to manually create our dataset from remixes of songs and modify the existing MedleyDB dataset for our purposes. We have chosen Demucs architecture as a basis for our neural network. We trained it from scratch to separate audio files into five distinct recordings containing drums, bass, vocals, guitars, and other accompaniment. We trained five models on MetaCentrum, which we evaluated objectively and subjectively. The implemented application serves as both a music player and an educational tool. The main feature is to allow users to listen to isolated instruments, for example, a guitar, and therefore more easily learn songs by ear. The application was subjected to user testing, and the knowledge learned will be used in future development.
|
|
|
Music Source Separation
Holík, Viliam ; Veselý, Karel (oponent) ; Mošner, Ladislav (vedoucí práce)
Neural networks are used for the problem of music source separation from recordings. One such network is Conv-TasNet. The aim of the work is to experiment with the already existing implementation of this network for the purpose of potential improvement. The models were trained on the MUSDB18 dataset. It was successively experimented with the change of the network structure, transforming signals from the time domain to the frequency domain for the purpose of calculating the loss function, replacing different loss functions with the original one, finding the optimal learning rate for each loss function and gradually decreasing the learning rate during the learning process. The best experiments according to the SDR metric were training with loss functions L1 and logarithmic L2 in the time domain with a higher initial learning rate with its gradual decrease during the learning process. In a relative comparison of the best models to the baseline, it is more than 2.5% improvement.
|
host ::
RSS.