Learning Target Template for Acoustic Event Detection From Low-SNR Training Data

Acoustic events of interest sometimes are inherently intermixed with strong background noise, like underwater sounds or mechanical fault signals. The noise usually creates a low-SNR scenario, making it more difficult to detect the events of interest. Suffering from lack of prior knowledge, it is a difficult job to extract "correct" features from noisy acoustic mixtures for the target events. This paper addresses the detection problem for acoustic events based on low-SNR training data. A novel learning algorithm is proposed to extract the target's template from noisy training data. Based on a target-guided band-pass filter using wavelet packets, which has been proposed in our previous work for low-SNR event detection, the template of a specific event is represented in the context of the wavelet packet energy spectrum. Template learning is then formulated as an optimization problem maximizing the detection performance which is quantified as the output SNR of the filtered signal. It is further proved that the original complex optimization problem is equivalent to a linear program which is easy to solve. A relatively sufficient database was built to simulate the noise-polluted training signals and the low-SNR detection task. Results show that the learning method can work at -15 dB and above, and the trained templates exhibit excellent detection performance comparable to the templates extracted from clean samples.