Discriminating speech and non-speech with regularized least squares

We consider the task of discriminating speech and non-speec h in noisy environments. Previously, Mesgarani et. al [1] achie ved state-of-the-art performance using a cortical representa tio of sound in conjunction with a feature reduction algorithm andonlinear support vector machine classifier. In the present wor k, we show that we can achieve the same or better accuracy by using a linear regularized least squares classifier directly on th e highdimensional cortical representation; the new system is sub tantially simpler conceptually and computationally. We selec t the regularization constant automatically, yielding a parameter -fr e learning system. Intriguingly, we find that optimal classifiers fo r noisy data can be trained on clean data using heavy regularization .