平成28（2016）年度　一般研究２実施報告書

課題番号	28−共研−2022		分野分類			統計数理研究所内分野分類			c
						主要研究分野分類			2
研究課題名	Music information processing using Deep Learning methods
フリガナ 代表者氏名	マルコフ コンスタンティン ペトロフ Markov Konstantin Petrov				ローマ字		Markov Konstantin Petrov
所属機関	会津大学
所属部局	情報システム学部門
職　　名	准教授
配分経費	研究費	40千円		旅　費		36千円		研究参加者数		2　人

研究目的と成果（経過）の概要

This project investigates several types of deep RNNs for dynamic music
emotion recognition (MER). Dimensional MER requires subject to annotate
the numerical VA values in the subjective test. It is used relationship
between some features xi and ground truth emotion values yi in MER. Then
data is split into Train data and Test data. In the train phase we train
regressor model for arousal and valence. We use SVR, FNN, LSTM as a
regressor model. These models predict arousal and valence of test data.
Last, evaluate each regressor. SVR is dealed as a baseline and compare
with models based on DNN.
The results of comparing performance of SVR with one of model based on
DNN. As a result, models based on DNN have a good performance than SVR.
Particularly FNN is acceptable performance. Surprisingly the best
performance was given in a small and narrow network with the hidden
node is 100 and the hidden layer is 2. We found that we are able to
achieve good performance even with a network much smaller than
we usually think in this experiment. Also, the performance of LSTM
became lower than FNN. That was unsuspected. The reason for this it
seems that it may be due to a mall amount of data with relationships
between data. We obtained better performance than the network alone by
combining different kinds of networks(in Valence). From this
results we would like to research the influence on the network to be
joined.

当該研究に関する情報源（論文発表、学会発表、プレプリント、ホームページ等）

No publications were submitted.

研究会を開催した場合は、テーマ・日時・場所・参加者数を記入してください。

No meetings were held

研究参加者一覧
氏名	所属機関
松井 知子	統計数理研究所