Sound Signal Processing Based on Seq2Tree Network

Most state-of-the-art solutions to sound signal processing tasks such as the speech and noise separation task and the music style classification task are based on Recurrent Neural Network (RNN) architecture or Hidden Markov Model (HMM). Both RNN and HMM assume that the input is chain-structured so that each element in the chain is equally dependent on all its previous units. However in real-life scenes the units alone do not carry much meaning. Only when several units group to be segments will they be semantically informative. This characteristic of sound signals clearly prefers emphasizing dependencies among units in the same segment, which leads to a natural selection of tree-structured models instead of chain-structured ones. In this paper we introduce Seq2Tree network and two models based on Seq2Tree architecture solving 1) speech and noise separation task and 2) music style classification task, respectively. Experiments show that our Seq2Tree-based models outperform the state-of-the-art systems in both tasks, which agrees with our hypothesis that sound signals have potential tree-structured dependencies among their sound elements. Also the experiment results prove the advancement of the Seq2Tree network architecture in sound signal processing tasks.