Tensorial dynamic time warping with articulation index representation for efficient audio-template learning.

Area of Science:

• Machine Learning
• Bioacoustics
• Signal Processing

Background:

• Audio classification typically requires extensive labeled datasets, which are difficult to obtain for applications like wildlife monitoring.
• Data scarcity and corruption (noise, interference) pose significant challenges in real-world audio classification tasks.

Purpose of the Study:

• To develop an efficient technique for learning robust audio templates from minimal, potentially corrupted labeled data.
• To improve the performance of template-based audio classification in data-scarce environments.

Main Methods:

• Utilizes tensorial dynamic time warping on articulation index-based time-frequency representations.
• Learns a clean template from a few labeled, noisy audio samples.
• Applies the learned template in a standard template-based audio classification framework.

Main Results:

• The proposed method successfully learns effective templates even with limited and noisy training data.
• Demonstrates superior performance compared to recurrent neural network (RNN) and state-of-the-art template-based methods.
• Achieves high accuracy in a wildlife detection application with few training samples.

Conclusions:

• The proposed template learning technique offers an efficient and effective solution for audio classification with limited labeled data.
• This approach is particularly beneficial for domains like wildlife monitoring where data acquisition is challenging.
• It provides a viable alternative to data-hungry deep learning models when training data is scarce.