A data-adaptive sound source separation system is presented, which is able to extract meaningful sources from polyphonic real-world music signals. The system is based on the assumption of non-negative sparse sources which have constant spectra with time-varying gain. Temporal continuity objective is proposed as an improvement to the existing techniques. The objective increases the robustness of estimation and perceptual quality of synthesized signals. An algorithm is presented for the estimation of sources. Quantitative results are shown for a drum transcription application, which is able to transcribe 66% of the bass and snare drum hits from synthesized MIDI signals. Separation demonstrations for polyphonic real-world music signals can be found at http://www.cs.tut.fi/~tuomasv/demopage.html.