ホットスポット 1 音リアクションイベント BIC GMM 2 3 BIC GMM HMM 10) SVM 11) 12) 13) Bayesian Information Criterion BIC 14) BIC M = M 1, M 2,,

Size: px

Start display at page:

Download "ホットスポット 1 音リアクションイベント BIC GMM 2 3 BIC GMM HMM 10) SVM 11) 12) 13) Bayesian Information Criterion BIC 14) BIC M = M 1, M 2,,"

さあしゃはしかわ
5 years ago
Views:

1 BIC GMM Acoustic Event Detection for Finding Hot Spots in Podcasts Kouhei Sumi, 1 Tatsuya Kawahara, 1 Jun Ogata 2 and Masataka Goto 2 This paper presents a method to detect acoustic events that can be used to find hot spots in podcast programs. We focus on meaningful non-verbal audible reactions which suggest hot spots such as laughter and reactive tokens. In order to detect this kind of short events and segment the counterpart utterances, we need accurate audio segmentation and classification, dealing with various recording environments and background music. Thus, we propose a method for automatically estimating and switching penalty weights for the BIC-based segmentation depending on background environments. Experimental results show significant improvement in detection accuracy by the proposed method compared to when using a constant penalty weight. 1. MP3 Podscope 1) PodCastle 2) Google Audio Indexing 3) PodCastle 4) 7) Google Audio Indexing 8) 9) 1) 1 Graduate School of Informatics, Kyoto University 2 National Institute of Advanced Industrial Science and Technology (AIST) 1 c 2009 Information Processing Society of Japan

2 ホットスポット 1 音リアクションイベント BIC GMM 2 3 BIC GMM HMM 10) SVM 11) 12) 13) Bayesian Information Criterion BIC 14) BIC M = M 1, M 2,, M m D = D 1, D 2,, D N M i BIC BIC(M i ) = log P (D 1, D 2,, D N M i ) 1 2 λd i log N (1) d i M i P M i BIC BIC 15),16) 1 N 1 M 0 = N(µ 0, Σ 0 ) BIC BIC(M 0 ) j 1 < j < N 2 M 12 = {N(µ 1, Σ 1), N(µ 2, Σ 2)} BIC BIC(M 12 ) BIC(j) = BIC(M 0 ) BIC(M 12 ) 2 c 2009 Information Processing Society of Japan

3 BIC(j) = 1 2 (N log Σ j log Σ1 (N j) log Σ2 ) 1 2 λ(d d(d + 1)) log N (2) d λ j = arg max j BIC(j) > 0 j λ 16) BIC GMM BIC λ λ 3. BIC GMM λ spe, λ mix, λ mus BIC 3.2 GMM GMM GMM GMM Nmix-GMM Gaussian 1 Gaussian m Gaussian N 十分大きな混合数分割分割分割 2 Gaussian 1-1 Gaussian 1-2 Gaussian m-1 Gaussian m-2 Gaussian N-1 Gaussian N-2 GMM GMM BIC 2 GMM BIC BIC= 1 2 ((ng m1 + n Gm2 ) log Σ Gm n Gm1 log Σ Gm1 n Gm2 log Σ Gm2 ) 1 2 λ m(d d(d + 1)) log(n G m1 + n Gm2 ) 0 (3) m = 1,, N n Gm1 n Gm2 EM m 1 m 2 3 m = 1,, N BIC 0 λ 3.3 λ 5 λ spe, λ mus, λ mix c 2009 Information Processing Society of Japan

4 GMM のパラメータ推定学習データ BIC の分割重み推定有声休止検出学習フェーズあいづち検出各クラスの GMM 各クラスの分割重み音声認識結果 3 笑い声検出入力音源特徴抽出前処理 ( 大分類 ) BIC セグメンテーション GMM による識別音声音楽区間検出 ( ホットスポット候補区間の切り出し ) GMM EM 256 BIC 12 MFCC 12 MFCC 26 16kHz MFCC 25ms 10ms 1 JNAS 17) RWC-MDB 18) 4.2 BIC JNAS+RWC-MDB JNAS, IMADE 19), Web GMM GMM GMM GMM BIC ( 1 ) W min 100 ( 2 ) ( 3 ) ( 4 ) BIC GMM t res GMM θ res ) 3 20) 4 c 2009 Information Processing Society of Japan

5 提案手法 Lambda=1.0 Lambda=1.5 Lambda= (a) program-open (b) program-closed Measure R P F λ = λ = λ = Measure R P F λ = λ = λ = GMM 1 19 program-open 23 program-closed GMM ( 1 ) λ ( 2 ) λ = 1.0 ( 3 ) λ = 1.5 ( 4 ) λ = 2.0 λ R P F F F F = (1 + α2 )RP R + α 2 P α α = λ c 2009 Information Processing Society of Japan

6 3 70% GMM 6. BIC GMM % 1) Podscope: 2) PodCastle: 3) Google Audio Indexing: 4) PodCastle : 2.0 SLP-65-7 (2007). 5) Goto, M., Ogata, J. and Eto, K.: PodCastle: A Web 2.0 Approach to Speech Recognition Research, Proc. Interspeech, pp (2007). 6) PodCastle : Web2.0 SLP-65-8 (2007). 7) Ogata, J., Goto, M. and Eto, K.: Automatic Transcription for a Web 2.0 Service to Search Podcasts, Proc. Interspeech, pp (2007). 8) Alberti, C., Bacchiani, M., Bezman, A. et al.: An Audio Indexing System for Election Video Material, Proc. ICASSP, pp (2009). 9) SIG-SLUD-A (2008). 10) Zhou, X., Zhuang, X., Liu, M. et al.: HMM-Based Acoustic Event Detection with AdaBoost Feature Selection, Multimodal Technologies for Perception of Humans, pp (2008). 11) Temko, A. and Nadeu, C.: Classification of acoustic events using SVM-based clustering schemes, Pattern Recogn., Vol.39, No.4, pp (2006). 12) Knox, M. and Mirghafori, N.: Automatic Laughter Detection Using Neural Networks, Proc. Interspeech, pp (2007). 13) Laskowski, K.: Contrasting Emotion-bearing Laughter Types in Multiparticipant Vocal Activity Detection for Meetings, Proc. ICASSP, pp (2009). 14) Schwarz, G.: Estimating the Dimension of a Model, The Annals of Statistics, Vol.6, No.2, pp (1978). 15) Chen, S. and Gopalakrishnan, P.: Speaker, environment and channel change detection and clustering via the Bayesian Information Criterion, Proc. of DARPA Broadcast News Transcription and Understanding Workshop, pp (1998). 16) Tritschler, A. and Gopinath, R.: Improved speaker segmentation and segments clustering using the Bayesian Information Criterion, Proc. Eurospeech, pp (1999). 17) : JNAS Journal of the Acoustical Society of Japan (E), Vol.20, No.3, pp (1999). 18) Goto, M., Hashiguchi, H., Nishimura, T. et al.: RWC Music Database : Popular, Classical, and Jazz Music Databases, Proc. ISMIR, pp (2002). 19) Kawahara, T., Setoguchi, H., Takanashi, K. et al.: Multi-modal recording, analysis and indexing of poster sessions, Proc. Interspeech, pp (2008). 20) ( : ) Vol.83, No.11, pp (2000). 6 c 2009 Information Processing Society of Japan

IPSJ-JNL5212048

IPSJ-JNL5212048 Vol. 52 No. 12 3363 3373 (Dec. 2011) environment. Experimental results show a significant improvement in detection accuracy by the proposed method. Furthermore, subjective evaluations suggest that hot