Comparison of audio input representations on piano transcription using neural networks

초록

We compare the effect of multiple input representations on polyphonic piano music transcription based on neural networks. A state-of-the-art piano transcription neural network model, onsets and frames, is explored. We first provide detailed backgrounds of the piano transcription and input representations for the readers who are unfamiliar with this area. For comparing their effects, we consider four spectrograms; Mel-spectrogram, Linear-spectrogram, Log-spectrogram and constant-Q-transform with various hyper parameters. The effects of frequency bins, Short Time Fourier Transformation (STFT) window size and hop length on the four spectrograms are also examined. Our results show that Mel-spectrogram of 2,048 STFT window size, 512 frequency bins and 256 hop length yields the highest accuracy. We show that Mel-spectrogram is one of the most satisfactory input representations in general. Mel-spectrogram dominates other spectrograms and keeps a relatively high transcription accuracy even at the low resolutions in our experiments.

키워드