Memory Based Automatic Music Transcription System for Percussive Pitched Instruments

The target of our work dealt with the problem of extracting musical content or a symbolic representation of musical notes, commonly called musical score, from audio data of polyphonic music of percussive pitched instruments. We focuses on note events and their main characteristics: the onset (note attack instant) and the pitch (note name). Signal processing techniques based on the Constant-Q Transform (CQT) are used to create a time-frequency representation of the signal. The onset detection algorithm operates on a frame-by-frame basis and exploits a suitable time-frequency representation of the audio signal. The solution proposed consists of an onset detection algorithm based on Short-Time Fourier Transform (STFT), and a classification algorithm based on Support Vector Machine (SVM) to identify the note pitch. We introduce a memory based feature vector for classification. Moreover, to ascertain the effect of the memory, we evaluated the accuracy of the corresponding memoryless system. Finally, to validate our method, we present a collection of experiments using a wide number of musical pieces of heterogeneous styles, involving recordings of polyphonic music of three percussive pitched musical instruments.