Discriminative Model for Polyphonic Piano Transcription
Automatic music transcription (AMT) aims to convert a musical signal into a symbolic representation. The challenge is that music has concurrent notes that overlap in time and interact in frequency, which results in a complex signal that is difficult to decode into a sequence of letters. Polyphonic piano music is a particular case, as it contains 88 pitches that can be played simultaneously. Despite years of research, AMT has not yet reached a satisfactory level of performance.
In this paper, we propose a discriminative model for polyphonic piano transcription. It combines a note-based acoustic model and a multivariable MLM to improve the performance of a frame-level transcription algorithm. The model uses a constant Q Fourier transform to represent the input signal in time and frequency domains. It then performs a multivariable MLM to estimate the probability of each note in the given frame, where the outputs of the MLM are used as classifiers for predicting the next frame-level note.
We also use a deep learning approach to train the model. Specifically, the MLM is trained with a deep convolutional neural network to learn a representation of the signal. We apply this learned representation to the AMT task. We test the model on the LabROSA dataset with real piano recordings. The model achieves a frame-level transcription accuracy of 68%, comparable to previous approaches.
https://tartalover.net/nicole-pesce
Proposing a Discriminative Model for Polyphonic Piano Transcription
Previous methods for AMT include support vector machines (SVMs) and bidirectional recurrent neural networks (RNN). SVMs are trained on spectral features, and the classifier outputs are temporally constrained by hidden Markov models. RNNs are able to capture more long-term dependencies than SVMs, but they suffer from the exploding or vanishing gradient problem and struggle with the complexity of a multivariable MLM. In our experiment, we replace the RNN cells with LSTM cells, which are better at memorizing information in sequences. The resulting MLM performs better than previous RNN-based systems.
https://m.youtube.com/@tartalover2682/videos
There are two main approaches to AMT: frame-based and note-based. Frame-based methods estimate pitches and onsets separately. One example is the harmonic temporal structured clustering approach by Kameoka. Another is the supervised non-negative matrix factorization (NMF) by Cogliati and Cheng. Note-based methods directly estimate the notes without detecting the onsets and have the potential to be more accurate than frame-based approaches.
Music transcription, the process of converting audio signals into symbolic representations such as sheet music or MIDI files, has been a subject of intense research in the field of music information retrieval (MIR). Polyphonic piano transcription, in particular, presents a challenging task due to the simultaneous occurrence of multiple notes played on the piano. In recent years, deep learning techniques have shown promising results in various MIR tasks, including polyphonic transcription. However, designing an effective model that accurately transcribes polyphonic piano music remains a significant challenge.
However, they have the drawback that they need to determine the notes at each frame-level by maximizing the likelihood of a set of candidate signals. This process can be improved by applying a beam search algorithm that repairs the prediction error locally. This improves recall and F-measure and helps to reduce the accumulation of errors over time.
