Bioacoustic fundamental frequency estimation: a cross-species dataset and deep learning baseline

Area of Science:

• Bioacoustics and computational biology.
• Animal communication and bioacoustics.

Background:

• Fundamental frequency (F0) is crucial for analyzing animal vocalizations, but manual estimation is laborious and automation is complex.
• Existing automatic F0 estimation methods from speech and music have limited progress in bioacoustics.

Purpose of the Study:

• To address the limitations in automated F0 estimation for bioacoustics.
• To create a benchmark dataset and evaluate deep learning algorithms for F0 estimation in diverse vertebrate vocalizations.

Main Methods:

• Compiled a benchmark dataset of over 250,000 vocalizations from 14 taxa, covering a wide range of acoustic properties (infra- to ultrasound, harmonicity, non-linear phenomena).
• Tested various algorithms, including neural networks, for F0 estimation on the benchmark dataset.
• Developed spectral measurements to assess F0 quality and correlate it with algorithm performance.

Main Results:

• Neural networks show potential for F0 estimation, even for unseen taxa or with unlabeled training data.
• Proposed spectral measurements of F0 quality correlate well with estimation performance.
• Current algorithm performance is not universally satisfactory across all taxa, but deep learning shows promise.

Conclusions:

• Deep learning approaches offer a path towards a more generic and reliable bioacoustic F0 estimation tool.
• The benchmark dataset and F0 quality metrics will aid the bioacoustics community in analyzing vocalizations.
• Further development of deep learning models is needed for comprehensive F0 contour analysis in diverse animal sounds.