Abstract:

In this paper we propose a method of audio chord estimation. It does not rely on any machine learning technique, but shows good recognition quality compared to other known algorithms. We calculate a beat-synchronized spectrogram with high time and frequency resolution. It is then processed with an analogue of Prewitt filter used for edge detection in image processing to suppress non-harmonic spectral components. The sequence of chroma vectors obtained from spectrogram is smoothed using self-similarity matrix before the actual chord recognition. Chord templates used for recognition are binary-like, but have the tonic and the 5th note accented. The method is evaluated on the 13 Beatles albums.