Automatic Music Transcription Using Event Based Piano Transcription
Event Based Piano Transcription
Automatic music transcription (AMT) is a key part of the Music Information Retrieval field of research. It converts music in audio format into MIDI pianoroll which can then be played on a synthesizer or used as a starting point for sheet music. This is a crucial step in music understanding as it allows the extraction of symbolic representations that can be translated back into different forms of music.
AMT has many applications including musical education, analysis and creation. It enables the creation of much larger training datasets for generative models. It can also improve the quality of a performance by helping to capture some of the musical nuances such as vibrato and articulation that are difficult for human listeners to perceive.
The state-of-the-art AMT model uses CNNs and LSTMs to predict the onsets of each note, which are then conditionally predicted on frames within a bar and a beat. The result is a list of events for each bar and beat that includes the type of event (note, chord etc) and its location in time. The timing data is usually divided into four columns which represent bars, beats, divisions and ticks. Bars are a unit of musical time and beats are a user-defined subdivision of a bar. Divisions are a user-defined fraction of a beat and ticks are miniscule units that can be user-defined in increments of an eighth or asixteenth.
Automatic Music Transcription Using Event Based Piano Transcription
There are several reasons why the AMT models used in this work are different from previous ones. The onset detection algorithm is optimized for frame-by-frame prediction using a binary time-frequency representation and a bank of SVMs combined with a constant Q transform for feature extraction. The classifier is asynchronous and uses a semi-CRF output layer that is quadratic in complexity. The resulting system performs well on a frame-by-frame basis and is able to distinguish note onsets from noise.
The timbral characteristics of the piano are also used as an additional feature in this system to help differentiate overlapping notes. In addition, the tempo changes that can occur in musical performances are accounted for as an extra factor when determining note durations. This is important because tempo changes can cause inconsistent perceptions of the rhythm and thus result in inaccurate transcription.
The method is currently being applied to a variety of piano recordings and has shown promising results so far. However, there are still a few challenges to overcome, such as improving performance in offset detection and extending the approach to polyphonic transcription of other instruments with less stable pitch. These issues will be addressed in future work. Weixing Wei, Peilin Li and Yi Yu. Licensed under Creative Commons Attribution 4.0 International License.
