The Sing Search system prototype
allows a user to sing a melody into a microphone and provides a list
of the closest matching songs from a database of polyphonic music.
The system consists of a user interface that allows the capture of audio melody
queries, a simple transcriber that converts the melody to a pitch
sequence, and a back end that performs melody matching to produce
a ranked list of answers to the query. An indexer creates a melody
index consisting of melodies extracted from a collection of polyphonic
(more than one note at a time) music.
The transcriber converts the audio wave-form of the sung query
into a sequence of note transitions that can be used for matching
independent of the key of the query and songs in the database.
The current melody matcher searches the index
for the distinct 5-grams of
symbols representing pitch transitions that exist in the query
melody to retrieve potentially matching pieces in the database.
Pieces are ranked based on
the count of distinct common 5-grams --- a technique shown to
work well for symbolic melody queries against a polyphonic collection.
In this paper I will describe the technology behind the system,
practical considerations, the limitations and further research that
will be embarked upon using sung melody queries.
Authors: Alexandra L. Uitdenbogerd
Event: SF08: Search and Information Extraction from Audio Data Workshop