High Resolution Piano Transcription
The ability to accurately transcribe music into sheet music is of significant interest to musicians and audio engineers. However, current AMT systems are limited by their low accuracy and their inability to handle dynamic changes in music. They also cannot transcribe sustain pedals, which are an essential feature of pianos’ musical expression.
This work presents a new approach to high resolution piano transcription using an analytical algorithm that predicts the presence or absence of both notes and pedals in audio recordings of piano performances. It is based on the assumption that each note and pedal event has an individual presence probability and uses a time-varying likelihood model to estimate this probability.
Unlike previous transcription systems, which split audio recordings into audio frames and used discriminative models to classify the presence or absence of onsets and offsets frame-wise, this method computes and stores continuous probability values for each note and pedal events, making it possible to perform pedal and note detection as well as high-resolution piano transcription.
A New Approach to High Resolution Piano Transcription
It shows that the proposed model achieves state-of-the-art piano onset and velocity detection performance, and that a relatively small convolutional network is sufficient to achieve high-resolution piano transcription. The model is also capable of predicting the onsets and offsets of both sustained pedal and non-sustained pedal events, which are crucial for capturing the full range of expressiveness in piano music. In addition, it demonstrates that the AMT model can perform pedal detection and transcription in a zero-shot context without any training data.
The study suggests that the key to improving AMT accuracy is to reduce transcription time by reducing the frame hop size and by using a more compact representation for onsets and offsets, as opposed to a truncated target (e.g., first to third rows of Fig. 1). This approach is more computationally efficient than previous methods and provides much higher transcription resolution in terms of both onset and offset detection accuracy, even for short durations such as one to three frames.
While the results are still far behind the capabilities of a trained pianist, the AMT system could be useful in several applications. For example, beginner piano students can use the AMT to self check their performance as they practice a piece of music, and it can help them identify their mistakes. The AMT system can also be incorporated into more complex software to help music students with their practicing by highlighting errors that can be difficult for them to recognize themselves. Moreover, AMT can be a helpful tool for professional musicians to use in recording sessions and when reviewing their own music.
