@techreport{oai:ipsj.ixsq.nii.ac.jp:00224473,
 author = {北村, 悠 and 菊地, 晏南 and 齋藤, 大輔 and 峯松, 信明 and Yu, Kitamura and Anan, Kikuchi and Daisuke, Saito and Nobuaki, Minematsu},
 issue = {76},
 month = {Feb},
 note = {地声と裏声の識別手法などがこれまでに提案されているが，両者の中間的な発声と言われるミックスボイスを分析する研究は少ない．本研究では，地声，裏声，ミックスボイスの声区を対象として，既存の音響特徴量を用いて声区情報を定量化することを目的とする．定量化の方法として，音声を地声と裏声の 2 クラスに分類する識別器にミックスボイスを入力し，推定されるクラス事後確率を利用した．本稿では先行研究を踏まえ，地声と裏声の差異を表す特徴量として，基本周波数，第二倍音と基本波の振幅差，1 次のケプストラム係数，非周期性指標，歌声フォルマントを検討した．これらの特徴量から構成された地声と裏声の識別器の性能評価のため，交差検証を行った結果，単一母音単一歌唱者の場合，およそ 95 % 以上の正解率が得られた．この識別器にミックスボイスを入力し，声区情報の定量化を行った結果，SVM では地声と裏声の中間的な事後確率が得られ，地声らしさ，裏声らしさを定量化することができた．, Methods to distinguish between modal and falsetto have been proposed so far, but there are few studies analyzing mixed voices, which are intermediate ones of them. This study aims to quantify vocal register information by using existing acoustic features for modal, falsetto and mixed voice. Class posterior probabilities were used to quantify it by inputting mixed voices to a discriminator that classifies voices into two classes, modal and falsetto. In this paper, fundemental frequency, amplitude defference between second harmonic and fundemental wave, the first cepstral coefficient, aperiodicity and singer’s formant were stated as features indicating difference between modal and falsetto. Cross-validation was performed to evaluate the performance of the discriminator between modal and falsetto composed of these features. As a result, the correct answer rate was about 95 % or higher in the case of single-vowel and single-singer. Mixed voices were input to this dicriminator. Posterior probabilities were intermediate between modal and falsetto in the case of SVM, and mixed voices were quantified between modal and falsetto.},
 title = {クラス事後確率に基づくミックスボイスを含めた声区情報の定量化},
 year = {2023}
}