Synthesis and Analysis Operators for Music Transcription

Abstract

Automatic music transcription (AMT) aims at obtaining a musical score from an input audio signal. AMT is particularly difficult when the audio signal is polyphonic, i.e., com- prises concurrent pitches such as chords. In this work we propose trainable sparse models for automatic polyphonic music transcription, which incorporates several successful ap- proaches into a unified optimization framework. The framework contains as particular cases standard sparse synthesis and analysis type of priors. Our models combines unsu- pervised synthesis models similar to latent component analysis and nonnegative matrix factorization with metric learning techniques that allow supervised discriminative learn- ing. The supervised training of the proposed model is formulated as a bilevel optimiza- tion problem, in which the prior parameters are optimized to achieve the best possible performance on the transcription task. We also show efficient approximation with fixed complexity and latency that can replace iterative minimization algorithms in time-critical applications.