Polyphonic Note Transcription of Time-Domain Audio Signal with Deep WaveNet Architecture

Deep neural networks-especially new, sophisticated architectures-are gaining increased competence at various tasks of growing complexity. Automatic transcription of polyphonic music is a complex problem with state-of-the-art approaches still suffering from high error rates relative to human-level performance. WaveNet is a novel deep neural network architecture with significant audio modelling capacity, operating directly on samples of digital audio signal. We propose an approach to polyphonic note transcription using WaveNet for feature extraction from raw audio and recognition of note presence. We show, that WaveNet can be used to detect pitch in polyphonic audio texture, as opposed to most other approaches, which mostly depend on time-frequency representations. Model outputs are smoothed and truncated using one of two alternative thresholding techniques. We evaluate our method on recently released dataset MusicNet and compare frame-level transcription performance to benchmark models. Results are promising, although further evaluation is needed to explore performance limits of this method.