Concurrent nonnegative matrix factorization using multi-resolution spectrograms for multipitch analysis of music signals

Bibliographic Information

Other Title
  • 時間周波数分解能の異なるスペクトログラムの並列NMFによる多重音解析

Search this article

Abstract

本報告では,自動採譜のためにスペクトログラムの非負値行列分解 (NMF) を用いた新しい多重音解析手法を提案する.自動採譜には音高の推定と発音時刻の推定が同時に必要であるが,音高と発音時刻の推定には解析フレーム長に関するトレードオフが存在する.そこで,異なる解析フレームによるスペクトログラムを併用することにより,双方の分解能を保ち音高と発音時刻の推定精度を高めることができると考えられることから,高時間分解能と高周波数分解能の 2 種類のスペクトログラムに対して NMF を並列的に用いる方法を提案し,実際の音響信号に対し簡単な発音検出実験を行い,従来の NMF と比較をしその有用性を示す.The Short-Time Fourier Transform (STFT) is commonly used as a back-end to multipitch analysis and it transforms acoustic signals into the time-frequency representation. However, a trade-off between time and frequency resolutions exists, depending on analysis frame length. It is therefore difficult to simultaneously obtain high accuracy of both note onset and note pitch estimation. Combination of spectrograms obtained with different frame lengths can achieve high resolution on both time and frequency, which should improve note parameter estimation. In this paper we propose a new method of multipitch analysis of musical signals based on Nonnegative Matrix Factorization (NMF) of spectrograms, where the NMF is applied in parallel to high time resolution and high frequency resolution spectrograms. We demonstrate the efficiency of our approach through note detection experiments.

Journal

Related Projects

See more

Details 詳細情報について

  • CRID
    1574231876753571328
  • NII Article ID
    110008584095
  • NII Book ID
    AN10438388
  • Text Lang
    ja
  • Data Source
    • CiNii Articles
    • KAKEN

Report a problem

Back to top