Transcription of the Singing Melody in Polyphonic Music

This paper proposes a method for the automatic transcription of singing melodies in polyphonic music. The method is based on multiple-F0 estimation followed by acoustic and musicological modeling. The acoustic model consists of separate models for singing notes and for no-melody segments. The musicological model uses key estimation and note bigrams to determine the transition probabilities between notes. Viterbi decoding produces a sequence of notes and rests as a transcription of the singing melody. The performance of the method is evaluated using the RWC popular music database for which the recall rate was 63% and precision rate 46%. A significant improvement was achieved compared to a baseline method fromMIREX05 evaluations.