Sardb: A Dataset For Audio Scene Source Counting And Analysis

Determining the number of sources in a signal is an important consideration for many audio scene analysis tasks. However, source counting is not actively researched like many other audio tasks. This work looks to create Ryerson University's Signal Analysis Research (SAR) group's SARdB: a multimodal audio-text dataset with the goal of promoting research on source counting and audio scene analysis. SARdB consists of 10s long acoustic scenes containing between 1 and 4 speakers and 0-5 sound events present for a total of similar to 21 hours of data. We demonstrate the utility in performing source counting and how it can be a benefit to audio scene analysis tasks in general. Crown Copyright (C) 2021 Published by Elsevier Ltd. All rights reserved.