A Probabilistic Subspace Model for Multi-instrument Polyphonic Transcription

In this paper we present a general probabilistic model suitable for transcribing single-channel audio recordings containing multiple polyphonic sources. Our system requires no prior knowledge of the instruments in the mixture, although it can benefit from this information if available. In contrast to many existing polyphonic transcription systems, our approach explicitly models the individual instruments and is thereby able to assign detected notes to their respective sources. We use a set of training instruments to learn a model space which is then used during transcription to constrain the properties of models fit to the target mixture. In addition, we encourage model sparsity using a simple approach related to tempering. We evaluate our method on both recorded and synthesized two-instrument mixtures, obtaining average framelevel F-measures of up to 0.60 for synthesized audio and 0.53 for recorded audio. If knowledge of the instrument types in the mixture is available, we can increase these measures to 0.68 and 0.58, respectively, by initializing the model with parameters from similar instruments.