TweetyBERT: Automated parsing of birdsong through self-supervised machine learning.

Area of Science:

• Bioacoustics
• Artificial Intelligence
• Animal Communication

Background:

• Current methods for analyzing animal vocalizations rely heavily on human-labeled data, which is time-consuming and limits scalability.
• Parsing animal vocalizations in a fully unsupervised manner presents a significant challenge in bioacoustics research.

Purpose of the Study:

• To introduce TweetyBERT, a self-supervised transformer neural network designed for the unsupervised analysis of birdsong.
• To demonstrate TweetyBERT's ability to autonomously learn and identify communication units within animal vocalizations.

Main Methods:

• Developed TweetyBERT, a transformer neural network utilizing a self-supervised learning approach for audio analysis.
• Trained TweetyBERT to predict masked or hidden audio fragments without human supervision or labels.
• Applied TweetyBERT to analyze canary song to identify behavioral units like notes, syllables, and phrases.

Main Results:

• TweetyBERT successfully learned to parse canary song, autonomously identifying behavioral units such as notes, syllables, and phrases.
• The model captured intricate acoustic and temporal patterns within the birdsong.
• Demonstrated the potential for self-supervised learning in analyzing complex animal communication systems.

Conclusions:

• Self-supervised models like TweetyBERT can effectively analyze animal vocalizations without human-labeled data.
• This approach significantly accelerates the analysis of large, unlabeled datasets of animal communication.
• TweetyBERT offers a promising new direction for bioacoustic research and understanding animal communication.