Automatic Piano Transcription
Automatic piano transcription (APT) is a challenging task as it requires the transcriber to have knowledge of musical notation. Notation is the set of symbols used to represent a music piece on a sheet, such as onsets and ends of notes, rhythms, chords, and melodies. The most important requirement for a good APT model is its ability to recognize music scores and produce high-quality outputs, which are musically meaningful to people.
To meet this requirement, a number of different models have been developed, such as recurrent neural networks (RNNs), deep learning architectures, and various other machine learning algorithms. However, most of these models cannot achieve competitive re- sults on real-world labelled music recordings from diverse mu- sical genres.
One reason is that they cannot deal with the dynamic characteristics of a music piece, such as its tempo and dynamics. Another reason is that they cannot transcribe music with a large number of instruments, which are usually not presented in the labelled training data. Therefore, many existing supervised AMT models do not produce results that can match human experts’ performance.
https://tartalover.net/nicole-pesce
Deep Learning in Automatic Piano Transcription
The underlying problem of these models is that they need to learn the whole note definition domain of a song in order to perform well, which is a very time-consuming process. Furthermore, they often fail to recognise the correct intervals or chords in a melody and do not accurately predict the onsets and offsets of each note.
https://m.youtube.com/@tartalover2682/videos
The aim of this research is to improve the accuracy and stability of piano music score recognition by introducing an attention mechanism based on a deep recognition algorithm. A composing model is also built, which can learn to compose music from scratch by using the attention mechanism, and it can produce more accurate prediction than traditional models.
In addition, this research aims to reduce the complexity of the composing model by removing the harmonic layer from it. This will enable the composing model to more easily overfit on the training data, and therefore it can be trained with less data than other models that use harmonic layers.
To do this, a variational autoencoder (VAE) is used to encode the input music sequence into its latent space, which is then decoded by an RNN to generate the desired musical output. The generative process is optimized by ensuring that the points sampled from latent space have good properties, such as continuity and completeness, to give music that sounds natural and has the appropriate amount of variation.
The VAE-NN method was found to be effective in generating music with the characteristics of a piano performance, including an attack stage with strong intensity, and a decay stage with decreasing intensity over time. The resulting music sequences were tested on the MAPS dataset and achieved competitive re- sults. The MAPS dataset is composed of nine parts, each containing 30 piano performances randomly selected from 238 different classical and traditional pieces. Each part is further divided into a test and train set, which are compared with each other to balance the dataset.