A non-negative approach to semi-supervised separation of speech from noise with the use of temporal dynamics