Overview

This poster summarizes ongoing work within the Soundscape group at MBARI toward characterizing humpback whale songs. We use some of the techiques widely applied for human speech analysis, in particular Linear Predictive Coding for signal processing and feature extraction, and Hidden Markov Modeling among other probabilistic models for supervised learning. Although we discuss the relative advantages of these methods in terms of classification accuracy, our main focus is on investigating the effect of some of the involved parameters in the processing pipeline, including order of prediction, analysis window size, and vector quantization codebook size. Overall, we are building a foundation for song structure analysis, toward understanding song as an aspect of humpback whale culture.

Goals

• Build upon signal processing and machine learning techniques and extend them into humpback whale song analysis

• Evaluate the effect of various signal processing parameters for song unit identification