Piano Transcription
Automatic music transcription (AMT) converts audio recordings of musical pieces into notation formats such as MIDI and sheet music. AMT has a wide range of applications from providing an easy way to play a piano piece on a synthesizer to creating training datasets for generative models that can emulate the style and interpretative choices made by human pianists.
The state-of-the-art AMT systems rely on a combination of CNNs and LSTMs to predict note onset events and condition framewise pitch predictions. This approach has achieved an excellent accuracy level, but is limited in its ability to capture the nuances of piano performance such as vibrato and articulation. To expand the capabilities of AMT to capture these expressive elements, this paper introduces an event based piano transcription framework that transforms audio into sequences of 0 and 1 events: note-on, note-off, time shift and velocity events.
A piano note is not a single frequency of sound vibration, but a composite of multiple frequencies with mathematically related harmonics. These harmonics can be detected in the time domain using zero-crossing techniques and in the frequency domain using techniques such as the discrete Fourier transform (DFT) or cepstrum. A number of approaches have been proposed to detect individual pitches, but none of them has been successful in transcribing polyphonic music.
Event Based Piano Transcription
To improve the quality of AMT, this paper proposes a new event-based approach that directly predicts note events using a semi-CRF output layer. This output layer is optimized to predict note onset events, while minimizing the prediction error in the surrounding frames. The algorithm also uses a new approach to detecting note offsets by regressing on the precise onset times and ignoring frames before or after them.
Piano transcriptions serve not only as vehicles for expanding the pianist’s repertoire but also as educational tools that offer insight into the compositional techniques and stylistic nuances of different musical genres. For students and performers alike, studying and performing piano transcriptions can provide valuable opportunities for musical growth and exploration.
One of the key benefits of studying piano transcriptions is the opportunity to delve into the intricacies of different musical styles and traditions. By transcribing and performing works originally written for other instruments or ensembles, pianists gain a deeper understanding of the unique characteristics and conventions of those genres.
For example, a pianist studying a transcription of a jazz standard can learn about swing rhythms, improvisational techniques, and the harmonic vocabulary commonly used in jazz music. Similarly, a classical pianist exploring transcriptions of Baroque or Romantic orchestral works can gain insight into the contrapuntal textures, expressive gestures, and formal structures typical of those periods.
To evaluate the performance of this framework, we apply it to a variety of challenging AMT tasks on the ISMIR 2014 transcripts dataset. The results show that our method achieves state-of-the-art accuracy in note-wise F-measure and AUC, surpassing the previous best. We also report a large improvement in precision, and a substantial increase in recall for all tasks. Moreover, our system is easily extensible to other instruments by changing only the output vocabulary of the network, while previously a different architecture would have been needed for each new task. Weixing Wei, Peilin Li and Yi Yu. Licensed under a Creative Commons Attribution 4.0 International License.
