The orca activity detection is a challenging task that prevails in underwater acoustics. Signal level discrimination of orca activity to that of noise signal is minimum, hence a topic of interest. The orca activity detection is a subtask of Computational Paralinguistics Challenge (ComParE) 2019. In this work, we study a few novel acoustic cues based on phase and long-term information to capture the artifacts from signal to detect orca activity. The phase of signal possesses definite signal characteristics which is completely random in case of noise signal. In this regard, we investigate instantaneous phase as an artifact for orca activity detection. Additionally, we believe that the long-term features can be more helpful to detect such artifacts than the conventional short-term acoustic features. We explore these two directions along with the state-of-the-art baselines on ComParE functionals, bag-of-audio-words and auDeep features for ComParE 2019. The studies reveal that the instantaneous phase as a single feature can perform better than the fusion of three baselines given as a benchmark for the challenge. Further, we perform a score level fusion of the acoustic features and the three baselines that further enhances the performance.
Cite as: Das, R.K., Li, H. (2019) Instantaneous Phase and Long-Term Acoustic Cues for Orca Activity Detection. Proc. Interspeech 2019, 2418-2422, doi: 10.21437/Interspeech.2019-1894
@inproceedings{das19b_interspeech,
author={Rohan Kumar Das and Haizhou Li},
title={{Instantaneous Phase and Long-Term Acoustic Cues for Orca Activity Detection}},
year=2019,
booktitle={Proc. Interspeech 2019},
pages={2418--2422},
doi={10.21437/Interspeech.2019-1894}
}