A hybrid Gaussian-HMM-deep-learning approach for automatic chord estimation with very large vocabulary

We propose a hybrid Gaussian-HMM-Deep-Learning approach for automatic chord estimation with very large chord vocabulary. The Gaussian-HMM part is similar to Chordino, which is used as a segmentation engine to divide input audio into note spectrogram segments. Two types of deep learning models are proposed to classify these segments into chord labels, which are then connected as chord sequences. Two sets of evaluations are conducted with two large chord vocabularies. The first evaluation is conducted in a recent MIREX standard way. Results show that our approach has obvious advantage over the state-of-the-art large-vocabulary-with-inversions supportable ACE system in terms of large vocabularies, although is outperformed by in small vocabularies. Through analyzing and deducing system behaviors behind the results, we see interesting chord confusion patterns made by different systems, which conceivably point to a demand of more balanced and consistent annotated datasets for training and testing. The second evaluation preliminarily demonstrates our approach’s superiority on a jazz chord vocabulary with 36 chord types, compared with a Chordino-like Gaussian-HMM baseline system with augmented vocabulary capacity.