We propose a discriminative model for polyphonic piano transcription. Spectral features are obtained individually for each note. To solve the overlapping partial problem, we apply spectral separation by estimating the spectral envelope for each note. For classifying purposes, support vector machines (SVM) are trained on the spectral energy inferred from these spectral features. We apply a scheme of one-versus-all (OVA) SVM classiﬁers to discriminate frame-level note instances. To decrease the high frequency notes residual energy due to the downward notes shared partials, a method to cancel the interferences from the downward notes to the upward notes has been developed. The classiﬁer output is ﬁltered with a hidden Markov model. Our approach has been tested with synthesized and real piano recordings obtaining very promising results.
Click to purchase paper as a non-member or login as an AES member. If your company or school subscribes to the E-Library then switch to the institutional version. If you are not an AES member and would like to subscribe to the E-Library then Join the AES!
This paper costs $33 for non-members and is free for AES members and E-Library subscribers.