Audio Chord Recognition with a Hybrid Recurrent Neural Network

In this paper, we present a novel architecture for audio chord estimation using a hybrid recurrent neural network. The architecture replaces hidden Markov models (HMMs) with recurrent neural network (RNN) based language models for modelling temporal dependencies between chords. We demonstrate the ability of feed forward deep neural networks (DNNs) to learn discriminative features directly from a time-frequency representation of the acoustic signal, eliminating the need for a complex feature extraction stage. For the hybrid RNN architecture, inference over the output variables of interest is performed using beam search. In addition to the hybrid model, we propose a modification to beam search using a hash table which yields improved results while reducing memory requirements by an order of magnitude, thus making the proposed model suitable for real-time applications. We evaluate our model's performance on a dataset with publicly available annotations and demonstrate that the performance is comparable to existing state of the art approaches for chord recognition.