Musical Instrument Recognition in Multi-Instrument Audio Contexts

Automatic musical instrument recognition is an important aspect of machine listening. In this project, we deal with instrument recognition in the multi-instrument audio contexts. We evaluate the performance of a traditional machine learning method in juxtaposition with a deep learning method in a supervised multi-label multi-output machine learning approach. We also tune a set of analysis parameters: {analysis window size, hop size, binarization threshold} to improve the performance. We investigate the possibility of improving the instrument recognition performance by using alternative data representations along with the original data. We consider two such sets of alternative data representations: 1) LRMS (left, right, mid, side) channel audio data derived from the stereo audio, and 2) The harmonic and residual representations derived from the original audio. We propose two different strategies to combine the models built using each of the data representation sets and evaluate their performance. Finally, we use the best combination strategy to merge the capabilities of individual models to improve the overall instrument recognition performance. With the shortlisted set of analysis parameters and the best combination strategy, we achieve an improvement of 14.25% in the macro f-score and 24.17% in the exact match ratio with respect to the baseline performance reported for our dataset