For a project I was asked to help with monophonic pitch detection (for piano), so in this repository I use and compare different algorithms and libraries to achieve the task.
I started out with aubio which has a nice variant of YIN called YinFFT. It is fast and >96% accurate on slow pieces (~1 note per second, which is realistic for this project). The regular YIN used in combination with a low pass filter gets almost the same accuracy, but is about 15x slower.
For onset detection aubio has several methods that perform well: simple local energy, high frequency content hfc and the popular specflux. After some tuning specflux performs best in my use case with ~95-97%.
The madmom library additionally provides superflux, which outperforms specflux by another 2%, but it is not optimized for real time detection, making it unnecessary slow. madmom also provides pre-trained Neural Nets, but they were not trained specifically for piano and in my tests they performed worse than superflux.