独立成分分析におけるSN比推定と成分置換是正による分離能力の違い

概要

論文の詳細を見る
かやのもり１１　再録論文In a real world environment where acoustic signals are contaminated with various noises, it is difficult to estimate the signal-to-noise (SN) ratio only from signals observed at microphones; The knowledge of acoustic transfer functions and original source signals is inevitable for SN ratio estimation. The present paper proposes a method to estimate the SN ratio approximately in the real world environment without the knowledge of transfer functions and source signals: the SN ratio is estimated after application of Independent Cotnponcnt Analysis (ICA) to the signals observed at microphones. The proposed method also works as a speech segment detector since detection of speech segments are necessarily carried out in the course of SN estimation. From several experimental results, the method has been confirmed to be valid. These results also show that although there have been some reports asserting that the separation ability of FDICA is equivalent to adaptive beam former (ABF), the assertion holds when the permutation of FDICA corrected by use of source location information, but does not hold when it is corrected by use of entropy.