Singing voice melody estimation is the task of calculating the fundamental frequency (f0) of the predominant voice in a piece of music containing multiple instruments. In this work, I evaluate the performance of several popular f0 estimation algorithms us-ing an annotated dataset (MedleyDB) of raw monophonic tracks, polyphonic mixes, and source-separated vocals. Many of the models were created to estimate the predominant melody and not necessarily the sung vocal melody, for example they could

more »

... capable of estimating the melody in instrumental music. Of the models tested, CREPE performs highly as a monophonic model, and Deep Salience and Encoder/Decoder perform highly as polyphonic models. By implementing source-separation as a preprocessing step, monophonic models such as CREPE, SPICE, and become viable options for the task of vocal melody estimation. These mono-phonic algorithms each perform signi˝cantly better in pitch accuracy on the source-separated vocal tracks compared to the polyphonic mixes. Additionally, each of the polyphonic algorithms tested increased in overall accuracy when using the source-separated tracks instead of the polyphonic mixes. I suggest further research im-plementing source-separation as a preprocessing step to vocal melody estimation using other source-separation tools and di˙erent datasets. Potential datasets could include tracks with overlapping vocal harmonies as well as di˙erent musical styles such as metal, rap, or non-western music.