High Resolution Piano Transcription With Pivots and Pedals
Piano Transcription With Pivots and Pedals
The piano is a unique instrument that combines the beauty of a stringed musical instrument with the complexity of a mechanical one. Unlike the electric guitar, which can be played with only a few fingers, the piano requires precise coordination of hands and feet to create a range of sounds from single notes to full chords. This complexity makes it especially challenging to automatically transcribe.
Recent neural network-based approaches to automatic music transcription (AMT) have achieved state-of-the-art performance on many tasks, including identifying onsets and offsets in audio recordings. However, these methods rely on a fixed number of frames to transcribe music, meaning that their transcription resolution is limited by the frame size. Additionally, these systems are sensitive to misaligned onset and offset labels in audio recordings.
This paper develops a new method for high resolution piano transcription with pivots and pedals, which addresses both of these limitations. In particular, we propose an analytical algorithm to predict the continuous onsets and offsets of all notes and pedals in audio recordings. This approach is more accurate than previous transcription methods, which simply classify the presence or absence of a note or pedal.
High Resolution Piano Transcription With Pivots and Pedals
Moreover, we introduce a more stable model for determining the transition time between onsets and velocities by annotating attack process signals in multiple consecutive frames. This improved model enables better identification of the transition between note states and results in more reliable performance on the AMT benchmark MAPS.
Finally, we use our new onset and velocity detection model to train a new piano transcriber that outperforms the current state-of-the-art on MAPS without requiring any complex post-processing. This improvement is mainly due to a better handling of the dynamics and the specific characteristics of the piano sound, especially at low velocities.
The resulting high-resolution piano transcript can be used to generate MIDI piano rolls, which can then be played on a synthesizer or used as the starting point for sheet music. In this way, the automatic transcription of a piano recording can open up new possibilities for analyzing music that isn’t easily available in notated form and creating much larger training datasets for generative models. In addition, it can also provide a more precise understanding of pianistic expression by allowing us to identify the moments when the player pivots or depresses the pedals.