MSC Thesis Project on machine learning methods for animal sound source separation
We have an open MSC Thesis Project on machine learning approaches to sound source separation problems in animal communication research.
The evolution of language is closely linked to sociality and the coordination of group-level actions. However, evaluating complex vocal communication strategies has been challenging as recordings of groups of animals contain overlapping vocalizations and sound separation methods have not yet been well established. Developing automated methods for signal separation and localisation would allow for continuous monitoring of acoustic exchanges in animal groups and the extraction of calls from each individual, creating a basis for future research regarding signal functionality, variability and social relevance. Recently, machine-learning approaches become feasible thanks to availability of large data sets that include ground-truth information about source location.
The goal of this project is to evaluate and compare diverse methods of sound source localization and test them on high-quality data sets of natural and artificial sound sources. Available sound signals are recorded either with hydrophones in-water or with microphones in-air, and location information is provided by GPS data or can be inferred from video recordings. In case of recordings of multiple sound sources, signals from animal-borne recording devices are available. The tasks of the student are to conduct a literature review of methods suitable for sound-source localization (and separation) and to benchmark their performance on natural and synthetic data sets of increasing complexity ranging from a single artificial source to multiple natural sources. The student will get access to massive data sets and to high-performance computing resources.
The successful student is expected to be well versed with signal processing and machine learning methods and has good programming and data science skills. The student will work in a team of engineers and researchers who develop behavioral monitoring systems and who research vocal communication in animals including monkeys, songbirds, and whales.
If you like challenging signal processing problems and would like to refine your multi-modal machine learning skills while learning about animal behavior, then this project might be for you. Keywords: beamforming, microphone array, accelerometer, animal monitoring, sensor fusion, deep learning, data augmentation, posture, tracking.
Jörg Rychen (jrychen (at) ini.ethz.ch), Kaja Wierucka (kaja.wierucka (at) gmail.com), and Richard Hahnloser (rich (at) ini.ethz.ch)