Transcription With Recurrent Neural Networks
Recent advances in polyphonic piano note transcription have been mainly achieved by using neural network architectures that detect different note states such as onset or sustain and model their temporal evolution. Most of them use separate neural networks for each of the aforementioned states, and optimize multiple loss functions.
In addition, they usually handle the temporal order of the note states by either imposing an ordering on the outputs or using a post-processing module. This paper investigates a unified approach to this problem, where multiple notes are predicted as softmax outputs in a single neural network with a single loss function and a temporal order is learned by an auto-regressive connection within the model.
First, a CNN is trained to detect the onset frames of a music signal. Then, another CNN is trained to estimate the pitches at those onset frames. The input of the pitch estimation CNN is a spectrogram slice centered at the onset frame detected by the previous CNN. The output of the pitch estimation CNN is a set of probabilities of 88 piano pitches at the detected onset frame. Those probability values are then used to form a matrix of softmax outputs, and the final result is the list of a possible 88 note segments.
Polyphonic Piano Note Transcription With Recurrent Neural Networks
Then the list of possible 88 note segments is analyzed to verify that each of them corresponds to a different pitch. A recurrent neural network is then used to train a model to translate the pitch to the corresponding string. Finally, the string is translated back to a spectrogram slice centered on one of the 88 notes. The spectrogram is processed with a Mel scaled transform, which provides better results than linearly scaled spectrograms.
We also compare the performance of our system with other state-of-the-art systems on the MAPS dataset by evaluating recall, precision and F-measure. The results show that our system achieves a high performance on note-level transcription, which is mainly due to the recurrent structure of the MLM.
Furthermore, the internet has enabled transcribers to connect with fellow musicians and enthusiasts from around the world, forming vibrant communities and networks of collaboration. Through online forums, social media groups, and virtual workshops, transcribers can share ideas, seek feedback, and support one another in their creative endeavors.
This paper introduces a new method to improve note-level polyphonic piano transcription by incorporating a MLM into the acoustic model. The MLM is trained to predict notes in a music sequence, and its recurrent structure enables it to capture the temporal dependencies between the notes. We also propose a note verification stage which reduces the number of false positive notes by verifying that the MLM predicts the correct notes for each of the blank onsets in the thresholding transcription results. The combination of the acoustic MLM with this note verification significantly increases the performance of the overall system. The performance of the MLM-based system is even better than that of the acoustic model alone, which indicates that MLMs are effective in note-level transcription.
