Audio Similarity Detection

Audio similarity detection plays a crucial role in various applications, such as music recommendations, speaker recognition, copyright infringement detection, and content-based retrieval. Various categories of similarity types such as Audio signal similarity, audio content similarity, audio perceptual similarity, speaker voice similarity, and audio caption similarity are covered by the researchers to identify similarity in audio signals. We have performed various experiments to investigate the appropriate audio embeddings for this task and to identify which feature and distance metrics are more appropriate for identifying speaker and content similarity. Various short-term spectral features have been extracted from the audio signal for feature extraction tasks which are then used for calculating a similarity matrix using Euclidean similarity and Cosine similarity. The performance of the fused features is better than the standalone features. Also, with the use of pre-trained features which are extracted from YAMNet, the model’s performance has been improved.