Filming the sound: Anomaly Detection on Audio Tape Recordings using Computer Vision Algorithms

This repository makes available the dataset related to the paper:

Zafer Çınar, Alessandro Russo, Matteo Spanio, Niccolò Pretto, and Sergio Canazza, Filming the Sound: Anomaly Detection on Audio Tape Recordings using Computer Vision Algorithms, IAI4CH, Bozen, 2024.

The dataset and the experiment are described in the publication above.

This repository contains two main directories (bold indicates directory names):

video samples: the actual videos used in the paper's experiment. This folder contains four subdirectories - 3.75 ips, 7.5 ips, 15 ips, and 30 ips - each representing a different playback speed (in inches per second). Within each subdirectory are several MP4 files, recorded on an A810 Studer open reel recorder, documenting the playback of magnetic audio tapes. The files follow the naming convention “Xips (Y).mp4,” where X represents the tape playback speed and Y is a serial number identifier for each video.

irregularities: the metadata for each video with timestamp and type of irregularity. The folder includes four CSV files - 3.75.csv, 7.5.csv, 15.csv, and 30.csv - corresponding to the playback speeds of the video samples. Each CSV file provides handmade annotations for its respective videos, with three columns:

video_id: name of the video file in the format “Xips (Y).mp4,” where X is the tape speed and Y is the ID number.

time_label: timestamp indicating the irregularity, formatted as HH:MM:SS.mls.

irregularity_type: category of the detected anomaly, which may be one of the following: “splice,” “shadow,” “end-of-tape,” or “annotation.”