Abstract—This paper presents a method for extracting vocal melodies from popular songs. Underlying the extraction proce-dure is a sinusoidal representation applied to the input song signal. The desired vocal melody is isolated by focusing on specific (amplitude- and frequency-modulated) sinusoids that are identified as vocal, with the identification based on minimum mean square error (MMSE) estimation of the singing voice. The experiment results show that sensitivity to vocality brings a 15% absolute gain in vocal pitch recall, and that the proposed system is effective in indexing a 95-song database in a query-by-singing application. I.