Digital signal processing approaches to bird song analysis
Citation:
O'REILLY, COLM JOSEPH, Digital signal processing approaches to bird song analysis, Trinity College Dublin.School of Engineering.ELECTRONIC AND ELECTRICAL ENGINEERING, 2017Download Item:
Abstract:
The ability to automatically analyze bird vocalizations would provide major assistance to zoologists in their behavioral and ecological studies. Revisions of taxonomy need to be made in cases where a population of species has evolved or diversified enough to be treated as a new unique species. Bird vocalizations are an important facet of the review process as song has a major influence in mate choice, hence why zoologists tend to study distinguishable songs when analyzing populations.
This thesis presents a method to quantify differences in bird vocalizations, inspired by dialect difference assessment from speech processing. The use of different codebooks is introduced. Reference to how easily pairs can be separated using Gaussian Mixture Model (GMM) classifiers with Mel-frequency Cepstral Coefficients (MFCCs) is also presented. As endangered populations will inevitably have small datasets, machine learning approaches may be hard to train. A forensics inspired pitch contour difference measure works with approximately 200 samples per population.
An improved pitch tracker for bird vocalizations, called YIN-bird, is also introduced. Carefully
tuned parameters improve pitch tracking with YIN (a commonly used pitch tracking tool in the speech community), but optimal parameters can change quickly even within one song. YIN-bird, a modified version of YIN which exploits spectrogram properties to automatically set a minimum Fundamental Frequency (F 0 ) parameter for YIN, is presented. Gross pitch errors on whistles and trills are reduced by up to 4% on a ground truth dataset of synthetic bird song with known pitch which is an improvement. This dataset was evaluated by expert listeners and described as ?sounding like original & can hardly tell it is synthetic?. A qualitative analysis of complex bird vocalizations is also presented.
Work then focuses on techniques to automatically extract acoustic features commonly used by zoologists. Signal processing techniques are employed to automate the extraction of the acoustic features: maximum, minimum and peak frequency, and bandwidth. YIN-bird and sine-tracking, a feature extraction method successfully applied to bird classification previously, are the automatic methods employed. The performance of automatic methods is compared to the manual method currently used by zoologists. Results are compared to a major study on species delimitation in the zoology domain. Both methods are well suited to this task, and demonstrate the strong potential to begin to automate the task of acoustic comparison of bird species.
The thesis then addresses the automatic segmentation of bird recordings into target and non-target regions using a speaker diarization approach. Standard approaches to automatic bird species identification are inspired by speaker verification methods, and tend to use all available bird vocalizations in a sample. Zoologists however will first tend to isolate important parts of a vocalization, discarding less discerning elements. Thus the automation of such segmentation work is highly desirable. The first stage is to segment target vocalizations from field recordings which is performed automatically in this thesis. Birds are typically recorded in their natural environment with many competing signals present. A speaker diarization approach to segmenting such recordings into target/non-target events is presented. This diarization approach uses Bayesian Information Criterion (BIC) Hierarchical Agglomerative Clustering (HAC) from the LIUM toolkit. MFCCs are used as features. The classification error is reduced from 19.20% to 15.01% when comparing the proposed system to a HMM-based segmentation as a baseline. A discussion of system performance and conditions which influence false alarms is given.
Sponsor
Grant Number
E3: The Engineering, Energy and Environment Institute
Trinity College Dublin (TCD)
Author's Homepage:
http://people.tcd.ie/oreilc16Description:
APPROVED
Author: O'REILLY, COLM JOSEPH
Advisor:
Harte, NaomiPublisher:
Trinity College Dublin. School of Engineering. Discipline of Electronic & Elect. EngineeringType of material:
ThesisCollections
Availability:
Full text availableMetadata
Show full item recordLicences: