Instrument Identification from Mixed Audio Signals

Abstract

Sound source separation has been an active research topic over the years. With the advent of deep learning, there has been many developments in this field. Some early works include the Independent Component Analysis(ICA), the Wave-UNet model with the advent of deep learning. Some recent works include the HTDemucs and Open- Unmix. Here, the work was done on the Open-Unmix architecture. The architecture involves spectrogram calculation using STFT, several Multi layer perceptron layers and three BiLSTM layers with skip connections. A modified form of this architecture was involved in this project where transformer was used. The result showed a slight increase in the SDR levels and reduced training time.