エネルギー変化の線形予測符号化に基づくリズム特徴量を用いた音楽印象識別

IPSJ

Bibliographic Information

Other Title
  • Rhythm Features Based on Linear Predictive Coding of Energy Variations for Musical Mood Classification

Search this article

Description

本論文では音楽の印象識別を高精度で行う特徴量として，線形予測符号化に基づくリズム特徴量（Rhythm feature based on Linear Predictive Coding：RLPC）を提案する．RLPCは，音響信号のエネルギー変化に対して線形予測符号化を適用することにより求められるケプストラムであり，音楽におけるリズムの周期性をとらえることが可能である．7つの印象に対する音楽印象識別実験により，ジャンル分類ならびに印象分類における5種類の従来のリズム特徴量との比較を行った．実験結果より，RLPCを用いた場合の平均識別率は83.7%であり，従来のリズム特徴量を用いた場合より1.3ポイント高い識別率が得られた．さらに，音量・音色・和音特徴量にRLPCを併用した場合の平均識別率は89.5%であり，音量・音色・和音特徴量のみを用いる場合と比較して2.0ポイント，従来のリズム特徴量を併用した場合よりも0.6ポイント高い識別率が得られた．また，各印象においてRLPCと従来のリズム特徴量で仮説検定を行った結果，4種類の従来のリズム特徴量に対して，RLPCを用いた場合の識別精度が有意であった．

In this paper, we propose a novel rhythm feature, which we call Rhythm feature based on Linear Predictive Coding (RLPC), to improve mood classification performance. The proposed feature is extracted with Linear Predictive Coding (LPC) on energy variations of an audio signal and is able to represent periodicity of rhythm in musical audio signals. To evaluate the proposed feature in comparison with 5 conventional rhythm features, mood classification experiments were conducted for 7 moods. From these experimental results, average accuracy of the proposed feature was 83.7% and 1.1 point higher than that of the conventional features. In addition, in case of combining base features, which indicate intensity, timbre, and harmony features, with RLPC, average accuracy was 89.5%. The accuracy was 2.0 point higher than that of base features and 0.6 point higher than that of combining base features with the conventional features. From results of hypothesis test on each mood, accuracies of the proposed feature were significant against that of the 4 conventional features.

Journal

Related Projects

See more

Keywords

Details 詳細情報について

Report a problem

Back to top