Evaluation of Enhanced F0-Trajectories for Speech Detection and Classification in Acoustic Monitoring

We evaluate the performance of enhanced F0-features for robustly detecting speech segments in noisy acoustic monitoring recordings. F0-features are extracted from the spectrogram based on the recently introduced shift autocorrelation (shift-ACF) and subsequent trajectory extraction. Speech detection is performed in a two-stage approach, involving both a classification and a segment extraction stage. We systematically evaluate the shift-ACF features and the speech detection performance using (i) purely synthetically generated data, (ii) a mix of synthetic speech and real noise background, and (iii) real speech and real noise background. In reviewing their strengths and weaknesses it turns out that shift-ACF based F0-features outperform classical features in several scenarios.