To implement the classifiers with the best accuracy, first we had to acquire a lot of samples.
We recorded more than 16 notes of two chromatic piano scales, from C4 to C6. Every note had approximately the same duration in time and were separate from each other, so that we could split the signal in time domain and extract the main frequency and the first harmonic of each note.
We processed each note recording with our algorithm and created an excel file that contains two types of data.
The first column contains the main frequencies of all the notes. The second column corresponds to the frequency of the first harmonic of each of the notes. Finally, the third column identifies each note with its name and scale, i.e. 440 Hz, 880 Hz, 'a4'.
As the training data are all obtained from piano recordings. the note detection system works better for piano recordings.