The Effect of Audio Degradation on Onset Detection Systems

The Effect of Audio Degradation on Onset Detection Systems

Peer-reviewed article not indexed in WoS or Scopus

Although many articles in the field of Music Information Retrieval have been introduced to improve onset detection systems, only the bare minimum focus on the degradation of input audio to increase detection accuracy. This article evaluates the accuracy of five onset detectors, including stateof-the-art machine and non-machine learning-based systems, and compares the influence of various types of audio signal degradation on musical onset detection. We used three different degradations based on impulse responses, Teager–Kaiser energy operator, and two MP3 compression settings. The results suggest that if MP3 compression of any settings is applied, the accuracy of detection systems is very similar. Using the energy operator as degradation has not improved overall detection but may offer the potential of pre-processing the neural network input signal for easier identification of onsets in a training phase. Furthermore, radio broadcast degradation increases the number of all predicted onsets in general, both true and false positives, resulting in better recall but worse precision. This information could be used to modify the pre-processing phase of neural network-based detectors and to optimize the sensitivity trade-off.

Although many articles in the field of Music Information Retrieval have been introduced to improve onset detection systems, only the bare minimum focus on the degradation of input audio to increase detection accuracy. This article evaluates the accuracy of five onset detectors, including stateof-the-art machine and non-machine learning-based systems, and compares the influence of various types of audio signal degradation on musical onset detection. We used three different degradations based on impulse responses, Teager–Kaiser energy operator, and two MP3 compression settings. The results suggest that if MP3 compression of any settings is applied, the accuracy of detection systems is very similar. Using the energy operator as degradation has not improved overall detection but may offer the potential of pre-processing the neural network input signal for easier identification of onsets in a training phase. Furthermore, radio broadcast degradation increases the number of all predicted onsets in general, both true and false positives, resulting in better recall but worse precision. This information could be used to modify the pre-processing phase of neural network-based detectors and to optimize the sensitivity trade-off.

audio; degradation; evaluation; music information retrieval; onset detection

audio; degradation; evaluation; music information retrieval; onset detection

IŠTVÁNEK, M.

2022

07.03.2022

Elektrorevue

Brno

1213-1539

Elektrorevue - Internetový časopis (http://www.elektrorevue.cz)

24

1

Czech Republic

1

13

13

http://www.elektrorevue.cz/cz/clanky/zpracovani-signalu/0/the-effect-of-audio-degradation-on-onset-detection-systems/