Fig. 3

Dynamic behavior (a, b) and ROC curve (c) for two different features. The LTSD feature considers some temporal context whereas the SNR feature relies on a single frame. The ROC curve plots the average detection rate during speech intervals P _d(η) vs. the average false-alarm rate during noise-only intervals P _fa(η) for varying values of the threshold η. The dynamic behavior reflects the time dependency of detections relative to speech onsets and offsets for a fixed threshold η _10% chosen such that P _fa(η _10% )=0.1