Dataset Card for RepeatAudio

Audio datasets referenced in Class-Agnostic Audio Repetition Counting. Contains two main sections:

• RS, RSN and RVN: Synthetic datasets containing varying levels of noise. Uniformly 10 seconds long, with 0-8 repetition events contained in each sample.
• Clocks, Heartbeats and Dolphins: Real-world derived samples with variable length across mechanical, ecological and medical domains.
• Relevant code can be found in this repo.

Dataset Sources

Synthetic data was generated with two source datasets:

• FSD50K: Provided audio for repetition events. JSON files indicate source audio for each sample.
• MUSAN: Used as noise in background of samples.

Real-world sets were extracted from various sources:

• Clocks: Samples originate from the Clocks subset of BBC Sound Effects.
• Heartbeats: Training set of PhysioNet Challenge 2016.
• Dolphins: Click Trains folder from Song et al.