A DETERMINISTIC ANNEALING EM ALGORITHM FOR AUTOMATIC MUSIC TRANSCRIPTION
Tian Cheng, Simon Dixon, Matthias Mauch · Queen Mary Research Online (Queen Mary University of London) · 2013
In the past decade, non-negative matrix factorisation (NMF) and probabilistic latent component analysis (PLCA) have been used widely in automatic music transcription. Despite their successes, these methods only guarantee that the decomposition converges to a local minimum in the cost function. In order to find better local minima, we propose to extend an existing PLCA-based transcription method with the deterministic annealing EM (DAEM) algorithm. The PLCA update rules are modified by introducing a “temperature” parameter. At higher temperatures, general areas of the search space containing good solutions are found. As the temperature is gradually decreased, distinctions in the data are sharpened, resulting in a more finegrained optimisation at each successive temperature. This process reduces the dependence on the initialisation, which is otherwise a limitation of NMF and PLCA approaches. The method was tested on two standard multi-instrument transcription data sets (MIREX and Bach10). Experimental results show that the proposed method significantly outperforms a state-of-the-art reference method, according to both frame-based and note-based metrics. An additional analysis of instrument assignment results shows that instrument spectra are typically modelled as mixtures of templates from several instruments.