An Audio Front End for Query-by-Humming Systems

In this paper, the problem of processing audio signals is addressed in the context of query-by-humming systems. Since singing is naturally used as input, we aim to develop a front end dedicated to the symbolic translation of voice into a sequence of pitch and duration pairs. This operation is crucial for the effectiveness of searching for music by melodic similarity. In order to identify and segment a tune, well-known signal processing techniques are applied to the singing voice. After detecting pitch, a novel postprocessing stage is proposed to adjust the intonation of the user. A global refinement is based on a relative scale estimated out of the most frequent errors made by singers. Four rules are then employed to eliminate local errors. This front end has been tested with five subjects and four short tunes, detecting some 90% of right notes. Results have been compared to other approximation methods like rounding to the nearest absolute tone/interval and an example of adaptive movi...

• Type: Conference
• Year: 2001
• Where: ISMIR
• Authors: Emanuele Pollastri
• Keywords: Audio Signals, Duration Pairs, ISMIR 2001, Music, Well-known Signal Processing