Audio2Midi: a ResNet-18 based architecture for Accurate Audio-to-MIDI Translation

With the development of digital music production and audio processing technology, the demand for converting audio files to MIDI format is increasing day by day. MIDI files have become an important tool for music production and analysis due to their ability to accurately represent music information. Traditional conversion methods from Audio to Midi such as the methods based on frequency spectral analysis always suffered from the poor performances especially dealling with complex harmonies and multi-instrumental recording. In addition, the request of a large amount of calculating resources limits some methods’ feasibility in real-time application. In this paper, a novel method based on ResNet-18 to convert audio files including MP3, WAV, etc into MIDI files, a widely used file type in computer music technology for storing musical notes, timeframes, and excitations, is proposed and implemented. Compared with other methods, the method proposed in this paper demonstrates certain superiority and robustness. Specifically, in the POP909 dataset, the performance of this method on R2 score(R-Square coefficient of determination), MSE(Mean Squared Error), and MAE(Mean Absolute Error) in the training set is 0.8809, 0.0274, and 0.0637. In addition, in the validation set and test set, the R2 score, MSE and MAE are 0.5823, 0.1010, 0.1370 and 0.5874, 0.1027, 0.1380, respectively.