Retrieving Sound Samples of Subjective Interest With User Interaction

Abstract

This paper concerns the retrieval of audio samples with a high degree of user interaction, motivated by a practical use case. We consider an open set recognition scenario in which the goal is to find all occurrences of a subjectively interesting sound selected by a user within a particular audio file. We use only a single starting example and maintain interaction through yes-no answers from the user, indicating whether any new retrieved sound matches the target pattern. We present a small dataset for this task and evaluate a baseline solution based on Nonnegative Matrix Factorization and greedy feature selection.

Searching For Loops And Sound Samples With Feature Learning

“…Below we describe all data and methods used in the study. The dataset has been previously used in [5], but here it is developed further with the separation of two distinct types of sounds. The method of active retrieval and the feature representation it uses are descrived in subsections B-D.…”

“…In this work, we focus on empirical evaluation of deep feature learning to the described retrieval scenario. We have previously published early results relying on the use of feature extraction techniques considered standard for music audio in this task [5]. We build upon the earlier work by extending the method through the use of deep feature learning and evaluate the result on an improved version of the dataset.…”