Automatic Recognition of Element Classes and Boundaries in the Birdsong with Variable Sequences

Area of Science:

• Bioacoustics
• Computational Biology
• Machine Learning

Background:

• Analyzing long-term sequential vocalizations, like birdsong, presents challenges for developmental and generational studies.
• Existing automatic speech recognition methods are suboptimal for biological applications due to the critical need for precise timing information.

Purpose of the Study:

• To develop automated systems for recognizing birdsong, a complex sequential vocalization.
• To address the specific properties of birdsong: note categorization, precise temporal structure, and probabilistic production rules.

Main Methods:

• A three-step recognition procedure: local classification, boundary detection, and global sequencing.
• Comparison of different arrangements of these steps.
• Development of a hybrid model using a deep convolutional neural network (CNN) and a hidden Markov model (HMM).

Main Results:

• A hybrid CNN-HMM model proved effective for birdsong recognition.
• Identified optimal method arrangements based on the need for accurate boundary detection.
• Introduced a novel measure to evaluate both note classification and boundary detection accuracy.

Conclusions:

• The proposed automated systems are effective for birdsong recognition.
• The methods can be adapted for sequential vocalizations in other species with similar properties.
• This work advances the automatic analysis of complex animal communication signals.