A Discriminative Model For Polyphonic Piano Transcription
Polyphonic Piano Transcription
Getting the most out of an audio recording is no small feat, and transcoding polyphonic music is no exception. Fortunately, there are many techniques available that have been tested in the real world. In particular, discriminative learning and artificial neural networks have been applied to low-level representations. In particular, it is possible to simulate the human intelligence required to recognize musical notation. It is also possible to build a program that can automatically infer musical notation from musical scores. Several recent advances in NLP tasks have motivated further investigations.
The first is the multi-pitch detection method, which is based on decomposing an input time-frequency representation into component activations. These activations are then mapped to narrowband spectra to support frequency modulation. The most impressive part of this method is the fact that each of these spectra is modelled as a discrete signal. Another trick of the trade is the use of a non-negative matrix factorization algorithm to achieve this feat. In fact, this is one of the few methods that can perform multi-pitch detection at a level of accuracy comparable to human transcribers.
The multi-pitch detection method is complemented by a discriminative model for polyphonic piano transcription. This model uses a recurrent neural network to estimate the prior probability of a sequence of note combinations. It can be incorporated into a time or metre based input representation, or interfaced with an external memory unit. The input time-frequency representation is an NMF. The input aforementioned is accompanied by a series of basis spectra, a clever technique that is not uncommon in audio and music transcription. These spectra are then shifted across log-frequency to facilitate frequency modulation.
A Discriminative Model For Polyphonic Piano Transcription
The best way to test the above system is to generate an appropriate test set. To this end, 54 test files over four splits have been used, yielding approximately 648, 000 frames at test time. This may not seem like a lot, but it is actually quite impressive given that a majority of the test files are disjoint instrument types. The test set was then analyzed for its merits in the context of the multi-pitch detection method. The results revealed that a single recurrent neural network was capable of transcribing polyphonic piano music accurately.
The best part of this system is that it can be applied to virtually any piano-like instrument. Moreover, its performance is remarkably comparable to that of a human transcriptionist, especially considering that most note events are reoccurring. The performance of the aforementioned system was also demonstrated using a real piano recording. It was found that the system is capable of transcribing music at a rate of around 68%.
The multi-pitch detection method was augmented by a recurrent neural network (RNN) based music language model. This model is a bit more sophisticated than the usual HMM. In particular, it incorporates a music-based temporal model to make the model more functional. In addition, it uses a musical score dataset to test its powers of persuasion.
