Detection of stop consonants in continuous noisy speech based on an extrapolation technique

In this contribution, we present an algorithm for stop consonant detection in continuous noisy speech at low signal-to-noise ratios (SNR) in the range of -5 to 7 dB. In our approach a signal extrapolation technique is used to predict the future samples based on present observations and then the predicted signal is used in the detection process. Experiments are performed over 700 utterances of 60 speakers taken from the TIMIT database with three different noise types. The detection performance achieved is 71% at -5 dB input SNR and 90.21% at 7 dB input SNR in case of white Gaussian noise at false alarm rates of 25.14% and 10.81% respectively. The detection performance results achieved are well in line with previous methods though our experiments are conducted at lower SNRs.