MUSIC SEARCH METHOD BASED ON NEURAL NETWORK

A music search method using a neural network includes the steps of: enabling a computer device to convert seed data which the computer device receives into a chromagram in a constant interval unit in a frequency domain; enabling the computer device to output a plurality of vector information represe...

Full description

Saved in:
Bibliographic Details
Main Author BAN, BYUNG HYUN
Format Patent
LanguageEnglish
Korean
Published 22.11.2017
Subjects
Online AccessGet full text

Cover

Loading…
More Information
Summary:A music search method using a neural network includes the steps of: enabling a computer device to convert seed data which the computer device receives into a chromagram in a constant interval unit in a frequency domain; enabling the computer device to output a plurality of vector information representing musical scales by inputting the data converted into the chromagram to a previously learned neural network; and enabling the computer device to search for specific music by performing sequence matching on a series of melodies represented by the plurality of vector information and a melody stored in a musical score database. The neural network outputs the melody by receiving a sound wave which is training data and previously performs reinforcement learning by a method for comparing the outputted melody with the musical score corresponding to the sound wave inputted to the neural network. Accordingly, the present invention can search for the same music if a musical component is identical. 신경망을 이용한 음악 검색 방법은 컴퓨터 장치가 입력받은 시드 데이터를 주파수 영역에서 일정한 구간 단위로 크로마그램(chromagram)으로 변환하는 단계, 상기 컴퓨터 장치가 크로마그램으로 변환된 데이터를 사전에 학습된 신경망에 입력하여 음계를 나타내는 복수의 벡터 정보를 출력하는 단계 및 상기 컴퓨터 장치가 상기 복수의 벡터 정보가 나타내는 일련의 멜로디와 악보 데이터베이스에 저장된 멜로디에 대한 시퀀스 매칭을 수행하여 특정 음악을 검색하는 단계를 포함한다. 신경망은 훈련 데이터인 음악의 음파를 입력받아 멜로디를 출력하고, 출력된 멜로디와 상기 신경망에 입력된 음파에 대응하는 악보를 비교하는 방식으로 사전에 강화학습을 수행한다.
Bibliography:Application Number: KR20170025908