@techreport{oai:ipsj.ixsq.nii.ac.jp:00211598,
 author = {舟木, 慶一 and Keiichi, Funaki},
 issue = {23},
 month = {Jun},
 note = {線形予測（Linear Predictive Coding:LPC）分析は，スマートフォンや Skype，LINE，ZOOM などに実装されている音声符号化などの音声処理において，最も成功をおさめている音声分析と言える．LPC は音声信号から低演算量で自己回帰 (AR) スペクトルを推定でき，その係数は LSP によるベクトル量子化 (VQ) により効率良く量子化できる．LPC の最も有用な利点は，逆フィルタにより残差信号を容易に算出できることである．これまでに，ARMA 分析，時変分析，複素分析，声帯音源推定など，様々な拡張がなされてきた．我々は，MMSE (Minimizing Means Square)，ELS (Extended Least Square)，LASSO などに基づく，解析信号から複素 AR パラメータを各サンプルで推定できる時変複素 AR（TV-CAR）音声分析を既に提案した．LASSO は，ℓ1 ノルム正則化法であり，多大な計算量が必要であるが，性能はそれほど高くない．本稿では，ℓ2 ノルム正則化 RLP，時間 RLP (TRLP)，およびそれらのハイブリッド基準に基づく TV-CAR 分析を提案し，IRAPT による ????0 推定を用いてその性能を評価する．, Linear Prediction (LP) is the most successful speech analysis in speech processing, including speech coding implemented on a smartphone, Skype, Line, ZOOM, or so on. The LP can estimate Auto-Regressive (AR) spectrum from speech signal with a small amount of computation, and its coefficients can be quantized by using Vector Quantization (VQ) with Line Spectrum Pair (LSP). The most vital point of the LP is to be able to estimate the residual signal easily using the inverse filter. Numerous improvements on the LP has been carried out, including ARMA analysis, time-varying analysis, complex analysis, glottal source estimation, or so on. We have already proposed time-varying complex AR (TV-CAR) speech analysis that can estimate complex AR parameters in any sample from an analytic signal based on MMSE (Minimizing Means Square), Extended Least Square (ELS), LASSO, or so on. The LASSO is an ℓ1-norm regularization method that requires massive computation with moderate performance. This paper proposed ℓ2-norm regularization based TV-CAR analysis, RLP, TRLP, and their hybrid method and evaluated the performance using ????0 estimation with the IRAPT.},
 title = {ℓ2ノルム正則化TV-CAR分析を用いた音声の????0推定},
 year = {2021}
}