Machine Learning-Based Analysis of Bird Vocalizations

by Burooj Ghani
Doctoral thesis
Date of Examination:2022-02-15
Date of issue:2022-04-01
Advisor:Prof. Dr. Florentin Wörgötter
Referee:Prof. Dr. Florentin Wörgötter
Referee:Prof. Dr. Sarah Hallerberg
Persistent Address: http://dx.doi.org/10.53846/goediss-9147

 

 

Files in this item

Name:dissertation_ghani.pdf
Size:11.2Mb
Format:PDF

The following license files are associated with this item:

Abstract

English

Acoustic signals are rich in information content. For humans the skill of listening to sounds and extracting relevant information comes effortlessly. However, for computers this task is quite challenging. Machine hearing entails developing computational methods to capture the approximate statistical structure of acoustic signals. The aim of this thesis is to build on the computational analysis tools that will allow for automated monitoring of bird species based on their vocalizations. We have relied on shallow classifiers that allows us to work in the realm of computationally inexpensive models that are not as data hungry. Such models can also be more suited for real-time bird species classification where hand held devices can be utilised to carry out classification. Apart from this, a random selection of both bird species and recordings has been employed to benchmark the classification performance for general-purpose multi-species classification. It is investigated in detail if and how classification results are dependent on the number of species and the selection of species in the subsets presented to the classifier. Furthermore, the analysis is extended to explore the intra-species differences in bird species vocalizations. The ornithologists have known that birds species vocalizations can vary even within the same species. Machine learning models are employed to map out (in geographical space) the vocal variation in widespread species in a way that does not require hundreds or thousands of hours of manual processing of recordings.
Keywords: Bioacoustics; Machine Learning; Audio Signal Processing
 
Statistik