IPSJ-SLP

Size: px

Start display at page:

Download "IPSJ-SLP"

こうしょあさま
6 years ago
Views:

1 F0 MFCC MFCCF % 90.2% A System for Automatic Discrimination between Singing and Speaking Voices on the Basis of Peak Interval of Spectral Change, F0, and MFCC Shimpei Aso, 1 Takeshi Saitou, 2 Masataka Goto, 3 Katsutoshi Itoyama, 1 Toru Takahashi, 1 Tetsuya Ogata 1 and Hiroshi G. Okuno 1 In this paper we describe a system that discriminates between singing and speaking voices. Given a clean speech signal, it outputs the likelihood of each of the singing and speaking voices. Previous systems use temporal transition of spectral envelope (MFCC) and fundamental frequency (F0) as discrimination features. Our system adds peak interval of spectral change as a phoneme duration feature and weights these features according to the duration of the input speech signal. Experimental results with one-second speech signal show that our system achieves 90.2 % accuracy compared to 86.7 % with previous systems. We also describe a real-time application demonstrating our system. 1. 1) 2) Mel-Frequency Cepstram Coefficients MFCCF % 2) Graduate School of Informatics, Kyoto University 2 School of Electrical and Computer Engineering, College of Science and Engineering, Kanazawa University 3 National Institute of Advanced Industrial Science and Technology (AIST) 1 c 2012 Information Processing Society of Japan

2 2),3) 2 ( 1 ) Klapuri 4) phenomenal accent ( 2 ) MFCC MFCC F0 MFCC 12 MFCC 12 F0 1 2) Random Splicing 5),6) ( 1 ) % ( 2 ) ( 3 ) 1 2 MFCC MFCC F0 2) GMM ˆd 1 g ad (x) = 1 N ˆd = N log f (x a(n); Λ ad ) (1) n argmax g ad (2) d=, x x a (n) x a n n N Λ ad d ( ) a GMM f x GMM Λ MFCC 12 MFCC 12 F % 2) 2 ( 1 ) c 2012 Information Processing Society of Japan

3 1 歌声データベース歌声 GMM パラメタ 1. 学習時 2. 識別 ( テスト ) 時特徴量抽出 EM アルゴリズム朗読音声データベース朗読音声 GMM パラメタ GMM 歌声 GMMパラメタテスト音声データ特徴量抽出尤度比較 GMM 朗読音声 GMM パラメタ ( 2 ) 3 F0 Random Splicing F0 F F0 3) F0 F0 Jitter 5 72 Jitter F0 7) STRAIGHT 8) F F0. F0 F0 2.1 F0 2.3 Klapuri 4) phenomenal accent 9) Klapuri 4) ( 1 ) 44.1kHz ( 2 ) 36 50Hz 20kHz 3 c 2012 Information Processing Society of Japan

2 * 4.1 ( 2 ) 3.2 ( 1 ) ( 2 ) ( 3 ) µ-law compression ( 4 ) z b (n) b n ( 5 ) z b(n) z b(n) = max(z b (n) z b (n 1), 0) ( 6 ) z b (n) z b(n) ( 7 ) 9) 2 3.

4 2 * 4.1 ( 2 ) 3.2 ( 1 ) ( 2 ) ( 3 ) µ-law compression ( 4 ) z b (n) b n ( 5 ) z b(n) z b(n) = max(z b (n) z b (n 1), 0) ( 6 ) z b (n) z b(n) ( 7 ) 9) ( 1 ) 4. F0 MFCC MFCC n t(n) t(n) t(n 1) 2 1 T t(n) t(n 1) > T P T, P T = 50[ ] P = c 2012 Information Processing Society of Japan

5 G a (x) = (g a (x) g a (x)) (3) x a g ad d F0 MFCC MFCC h ˆd h(x) = w a G a (x) (4) ˆd = { a F h(x a ) > 0 h(x a) 0 F F0 MFCC MFCC 4 w a x G a(x) M M 2 1 (5) 3 入力 GMM 歌声 GMMパラメタ出力音声データ特徴量抽出 GMM 朗読音声 GMM パラメタ歌声尤度話声尤度尤度差 1 F0 16kHz 10 1kHz 80Hz 4.4 単独特徴量識別器 1 4 単独特徴量識別器 2 入力音声入力音声長に応じた重みづけ和出力尤度差単独特徴量識別器 N. 2 MFCC 16kHz ,000Hz F0 yegnanarayana 10) F0 1 MFCC 2 F0 MFCC 2K + 1 K k c[n + k] k= K c[n] = K (6) K k2 c[n] n K c 2012 Information Processing Society of Japan

6 AIST 11) RWC : F0 MFCC MFCC 5 MFCC F0 MFCC % 74.8% MFCC 71.9% 78.3% F0 78.6% 84.5% MFCC 86.3% 91.8% 5.3 ( 1 ) F0 MFCC MFCC 3 ( 2 ) F0 MFCC MFCC 3 ( 3 ) F0 MFCC MFCC % 89.8% 89.5% 94.1% 90.2% 94.6% 3 F0 F0 MFCC MFCC MFCC MFCC 4 F0 MFCC MFCC ] [ % 度精別識 5 ] [ % 度精別識 F0 MFCC MFCC 評価音声長 [ ミリ秒 ] ピーク間隔 ΔF0 MFCC ΔMFCC F0 MFCC MFCC 評価音声長 [ ミリ秒 ] 大石手法ピーク間隔無し本統合法本統合法 ] [ % 度精別識数係み重評価音声長 [ ミリ秒 ] ピーク間隔 ΔF0 MFCC ΔMFCC 大石手法ピーク間隔無し本統合法本統合法評価音声長 [ ミリ秒 ] ピーク間隔 ΔF0 MFCC ΔMFCC 6 c 2012 Information Processing Society of Japan

7 MFCC 8 ( 1 ) MFCC MFCC ( 2 ) F0 MFCC MFCC F0 7. F0 MFCC MFCC 9 ( 1 ) Voice Activity Detection VAD ( 2 ) ( 3 ) ( 4 ) 2, VAD 9 入力音声 ΔMFCC 識別機 (1) VAD による波形切り出し (2) 単独特徴量に基づく識別 MFCC 識別機 ΔF0 識別機 (3) 音声長に応じた重み付け和アクセントピーク間隔識別機 (4) 表示モジュール話声らしい統合識別結果歌声らしい話声らしいアクセントピーク間隔歌声らしい話声らしい ΔF0 歌声らしい話声らしい MFCC 歌声らしい話声らしい ΔMFCC 歌声らしい 4 MFCC % 90.2% % 7 c 2012 Information Processing Society of Japan

8 NTT GCOE 1) WWW. SP, Vol.97, No.560, pp (1998). 2) Vol.47, No.6, pp (2006). 3) F0 73, 2R-6 (2011). 4) Klapuri, A., Eronen, A. and Astola, J.: Analysis of the meter of acoustic musical signals, Audio, Speech, and Language Processing, IEEE Transactions on, Vol.14, No.1, pp (2006). 5) Scherer, K. R., Feldstein, S., Bond, R. N. and Rosenthal, R.: Vocal cues to deception: A comparative channel approach, Journal of Psycholinguistic Research, Vol.14, pp (1985). 6) Friend, M. and Farrar, M.J.: A comparison of content-masking procedures for obtaining judgments of discrete affective states., J Acoust Soc Am, Vol.96, No.3, pp (1994). 7) Erickson, D., Suzuki, T., Tanosaki, K., Yahiro, K., Haneishi, E. and Kishimoto, H.: Ah, how sweet the sound: Some acoustic characteristics of emotionally sung /ah/, Intersinging (2010). 8) VOCODER:STRAIGHT Vol.54, No.7, pp (1998). 9) Ryynanen, M. and Klapuri, A.: Modelling of Note Events for Singing Transcription, ISCA Tutorial and Research Workshop on Statistical and Perceptual Audio (2004). 10) Yegnanarayana, B., Sri Rama Murty, K. and Rajendran, S.: Analysis of stop consonants in Indian languages using excitation source information in speech signal, ISCA ITRW Speech Analysis and Processing for Knowledge Discovery (2008). 11) AIST : Vol.2005, No.82, pp.7 12 (2005). 8 c 2012 Information Processing Society of Japan

スライド 1

スライド 1 スペクトル変化量のピーク間隔 F0 MFCC を用いた歌声と朗読音声の自動識別システム阿曽慎平 ( 京大 ), 齋藤毅 ( 金沢大 ), 後藤真孝 ( 産総研 ), 糸山克寿, 高橋徹, 尾形哲也, 奥乃博 ( 京大 ) サウンドテスト人 : 歌声話声 ( 朗読音声含む ) 聞き分け応答奥乃研音楽情報処理グループの日常天気悪いなぁ雨降るらしいでアイウォンチュ ~ アイウォンチュ ~