Detection of speech overlapped with low-energy music using pyknograms