The MIT used a carbon/electret microphone detector based on a GMM (Gaussian Mixture Model) of 1024 Gaussian. They used 5 hours of speech coming from LLHDB database to train their detector.