This paper presents a supervised algorithm for separating speech from background non-stationary noise (piano music) in single-channel recordings. The proposed algorithm, based on a nonnegative matrix factorization (NMF) approach, is able to extract speech sounds from isolated or chords piano sounds learning the set of spectral patterns generated by independent syllables and piano notes. Moreover, a sparsity constraint is used to improve the quality of the separated signals. Our proposal was tested using several audio mixtures composed of real-world piano recordings and Spanish speech showing promising results.