Audio similarity detection algorithm based on Siamese LSTM network

Abstract

The key technology of audio signal similarity detection lies in the selection of audio signal features and feature matching model. In order to improve the accuracy of the similarity calculation of audios, a method of using LSTM in the basic network part of the Siamese network is proposed. First of all, we extract the Filter banks features of the two audio signals. Then, two feature matrices are input into the network to calculate the audio similarity. Experiments show that the Siamese LSTM network using FBank features can accurately detect the similarity of two audio segments.