Soft frame margin estimation of Gaussian Mixture Models for speaker recognition with sparse training data