The assignment is to be able to recognize patterns in audio data. Our goal for this research was to identify musical instruments using Machine Learning algorithms. The intended value of the project is to get a better understanding how Machine Learning can effectively be used for these kinds of tasks.The previous group already made an algorithm that is able to recognize musical notes played by a piano. We decided to go a different route and try to recognize instruments.
One of the biggest challenges was to find a suitable dataset for our problem with enough quality data and good labels. Another challenge was to get a good understanding of which audio features to use, how the model is making decisions based on these features and how to determine the importance of each feature.
The expectations were that we could recognize instruments based on monophonic audio data. We also expected timbre to be a key factor in the recognition.
We trained a support vector machine (SVM) model and a simple neural network with two different datasets. One dataset with only monophonic audio data and the other dataset with melodies played by a single instrument. On the first dataset the SVM model showed an accuracy of 85% and the neural network an accuracy of 93%. Using the second dataset with the same method the SVM showed a accuracy of 94% and the neural network an accuracy of 99%. The difference in performance might be because of the different nature of the samples.
![](https://i.ytimg.com/vi/i-4ffro9mEU/mqdefault.jpg)