
Master thesis on Passive Audio Monitoring Systems
The present thesis examines the relationships between passive audio monitoring system (PAMS) algorithms for classification, localization and sound-shading objects, the sound alignment and the used in vitro bird sounds of three different bird species. In detail, four hypotheses are tested:
- The quality of the classification correlates positively with the quality of the localization
- The quality of the localization depends on:
- The sound shading of the sounder to the microphone,
- The kind of sound,
- The direction of the sound.
Figure 1: Map of the experimental setup in the courtyard of the campus.
For examination, a PAMS consisting of four Raspberry Pis were installed and calibrated. Stanzas of the bird species Alauda arvensis, Carduelis carduelis and Phylloscopus collybita were played at five equally measured speaker positions. The recording logic was programmed in Python using the pyalsaaudio package and the recordings were saved as WAV files with timestamps synchronized via NTP. For classification of the recorded data the spectrogram cross correlation of the R package monitoR was used. The location location was treated as an optimization problem and performed via the SciPy.minimize module. As a measure of the sound shading of signal position to the sensors, the laser scanning returns of the Digital Surface Model NRW were used.
In total, 15,412 bird sounds were recorded for the five signal positions. 14.8% of the Phylloscopus collybita, 15.5% of the Alauda arvensis and 100% of the Carduelis carduelis detections were wrongly classified. Carduelis carduelis detections were classified 81.5% as Phylloscopus collybita and 18.5% as Alauda arvensis. The localization showed deviations from the signal to the localized point of x 34.96 19.49m (min: 1.1, max: 95.8m), with clusters at 10 and 45m deviation at different times and at different play positions.
There was a very weak negative correlation between the quality of the classification and the quality of the localization (Pearson’s product-moment correlation -0.016, p-value 0.042). A significant negative correlation could be determined between the sound shading and the location accuracy (Pearson’s product-moment correlation -0.29, p-value <-2.2 ^ -16). The bird species played has a significant influence on the positional location quality (Kruskal-Wallis rank sum test, p-value 1.6 ^ -7). No significant corellation on the quality of the localization could be determined for the bird’s eye play direction (Pearson’s product-moment correlation -0.07, p-value <-2.2 ^ -16).
The complete thesis is available for download: http://eplus.uni-salzburg.at/obvusbhs/document/preview/2673790