An End-to-end Framework for Audio-to-Score Music Transcription on Monophonic Excerpts

In this work, we present an end-to-end framework for audio-to-score transcription. To the best of our knowledge, this is the first automatic music transcription approach which obtains directly a symbolic score from audio, instead of performing separate stages for piano-roll estimation (pitch detection and note tracking), meter detection or key estimation. The proposed method is based on a Convolutional Recurrent Neural Network architecture directly trained with pairs of spectrograms and their corresponding symbolic scores in Western notation. Unlike standard pitch estimation methods, the proposed architecture does not need the music symbols to be aligned with their audio frames thanks to a Connectionist Temporal Classification loss function. Training and evaluation were performed using a large dataset of short monophonic scores (incipits) from the RISM collection, that were synthesized to get the ground-truth data. Although there is still room for improvement, most musical symbols were correctly detected and the evaluation results validate the proposed approach. We believe that this end-to-end framework opens new avenues for automatic music transcription