Large vocabulary automatic chord estimation using bidirectional long short-term memory recurrent neural network with even chance training

This paper presents an argument for the necessity of a large vocabulary in automatic chord recognition systems, on the grounds of the requirements of machine musicianship. It proposes a system framework with a skewed class-sensitive training scheme that leads to a preliminary solution to large vocabulary automatic chord estimation. This framework applies a bidirectional long short-term memory recurrent neural network architecture, which employs an ‘even chance’ training scheme to make up for the lack of uncommon chords’ exposure. The main drawback of this approach is the low segmentation quality, which inevitably lowers the upper bound of chord estimation accuracy. Under a large vocabulary evaluation, the proposed system can significantly outperform the baseline system in terms of the overall weighted chord symbol recall, and there is no significant difference between them in terms of average chord quality accuracy. The results demonstrate preliminary success in our approach, and also prove the even chance training scheme to be effective in boosting uncommon chord symbol recalls as well as the average chord quality accuracy.