THE INSTITUTE OF ELECTRONICS, TECHNICAL REPORT OF IEICE. INFORMATION AND COMMUNICATION ENGINEERS

Size: px

Start display at page:

Download "THE INSTITUTE OF ELECTRONICS, TECHNICAL REPORT OF IEICE. INFORMATION AND COMMUNICATION ENGINEERS"

かずひろつくとの
5 years ago
Views:

116(477): 283-286 Issue Date 217-3 URL http://hdl.handle.

1 Title とメルケプストラムを用いた音響モデルに基づく騒音環境下叫び声検出の性能評価 Author(s) 福森, 隆寛 ; 中山, 雅人 ; 西浦, 敬信 ; 南條, 浩輝 Citation 電子情報通信学会技術研究報告 = IEICE technical re 信学技報 (217), 116(477): Issue Date URL RightCopyright 217 by IEICE Type Journal Article Textversion publisher Kyoto University

2 THE INSTITUTE OF ELECTRONICS, TECHNICAL REPORT OF IEICE. INFORMATION AND COMMUNICATION ENGINEERS Mel-Frequency Cepstrum Coefficients: MFCCMFCC MFCC Gaussian Mixture Model GMM Hidden Markov Model HMM Deep Neural Network DNN MFCC DNN GMM HMM Performance evaluation of noisy shouted detection based on acoustic with rahmonic and mel-frequency cepstrum coefficients Takahiro FUKUMORI, Masato NAKAYAMA, Takanobu NISHIURA, and Hiroaki NANJO College of Information Science and Engineering, Ritsumeikan University Nojihigashi, Kusatsu, Shiga, , Japan. Academic Center for Computing and Media Studies, Kyoto University. Nihonmatsu-cho, Yoshida, Sakyo-ku, Kyoto, , Japan. Abstract This paper describes a method based on new combined features with mel-frequency cepstrum coefficients (MFCCs) and rahmonic in order to robustly detect a shouted in noisy environments. MFCCs collectively make up mel-frequency cepstrum, and rahmonic shows a subharmonic of fundamental frequency in the cepstrum domain. In our previous method, Gaussian mixture s (GMM) is constructed with the proposed features extracted from training data which includes a lot of normal and shouted samples. In this paper, evaluation experiments of noisy shouted detection were conducted using not only GMM but also hidden Markov s (HMM) and deep neural network (DNN). The results show that MFCCs and rahmonic were effective for representing an utterance mechanism including both vocal tract and vocal cords. In addition, DNN could achieve higher performance in noisy environments than GMM and HMM. Key words Shouted detection, Noisy environment,, Mel-frequency cepstrum coefficients 1. [1], [2] This article is a technical report without peer review, and its polished and/or extended version may be published elsewhere. Copyright 217 by IEICE

3 [3] [5] [6] [9] SNR Mel-frequency cepstral coefficients: MFCC GMM Gaussian Mixture Model [1], [11] [12] GMM GMM HMM Hidden Markov Model DNN Deep Neural Network [12] [13] [14] [12] 1 2 1(a) 1(b) 2(a) 2(b) Power [db] Amplitude Frequency [khz] Power [db] Frequency [khz] (a) (b) 1 F Quefrency [ms] Normal Shouted Amplitude F Quefrency [ms] (a) (b) 2 [GMM, HMM or DNN construction] Recorded [Shouted detection] Observed 3 GMM, HMM or DNN Normal Shouted GMM, HMM or DNN Normal Shouted Result Normal or Shouted GMM Gaussian Mixture Model (GMM) Hidden Markov Model (HMM) Deep Neural Network DNN GMM HMM [15] HMM DNN

1 Training data Female speaker: 4 samples Male speaker: 4 samples Testing data Female speaker: 1 samples Male speaker: 1 samples Sampling 16kHz/16bit Acoustic 12 orders MFCC feature 12 orders ΔMFCC 1

4 1 Training data Female speaker: 4 samples Male speaker: 4 samples Testing data Female speaker: 1 samples Male speaker: 1 samples Sampling 16kHz/16bit Acoustic 12 orders MFCC feature 12 orders ΔMFCC 1 order Acoustic 1. GMM 2. HMM ( states) 3. DNN Noise White noise, Speech bubble [18] SNR, 1, 2, db Frame length 25 ms (Hamming window) Frame shift 1 ms DNN 3. Identification accuracy [%] Identification accuracy [%] MFCCs MFCCs+ GMM Mixed number 4 GMM MFCCs MFCCs+ HMM Mixed number 5 HMM 3. 1 ( 4 ) 4 SNR, 2, 1, db (GMM 3 HMM DNN GMM HMM GMM HMM HTK [16] DNN Kaldi [17] DNN NOISEX-92 [18] ([%] GMM HMM 4 GMM HMM 6 64 GMM HMM HMM 2 3 DNN 6 M, R, M+R DNN SNR 9 %

5 2 [%] ( DNN) Number of SNR= db SNR=2 db SNR=1 db SNR= db input frames M R M+R M R M+R M R M+R M R M+R 1frame frames frames *M: MFCCs, R:, M+R: MFCCs and 3 [%] ( DNN) Number of SNR= db SNR=2 db SNR=1 db SNR= db input frames M R M+R M R M+R M R M+R M R M+R 1frame frames frames *M: MFCCs, R:, M+R: MFCCs and DNN SNR= db GMM HMM 9 % DNN 95 % DNN GMM HMM DNN 4. GMM HMM DNN DNN SNR 9 % GMM HMM 16K1694 [1] W. Yao-Dong, T. Takeshi, and I. Idaku, HFR-video-based machinery surveillance for high-speed periodic operations, Journal of System Design and Dynamics, vol. 5, no. 6, pp , 211. [2] W. Huang, T. K. Chiew, H. Li, T. S. Kok, and J. Biswas, Scream detection for home applications, 5th IEEE Conference on Industrial Electronics and Applications, pp , 21. [3] M. Cowling, Comparison of techniques for environmental sound recognition, Pattern Recognition Letter, vol. 24, no. 15, pp , 23. [4] K. M. Kim, J. W. Jung, S. Y. Chun, and K. S. Park, Acoustic intruder detection system for home security, IEEE Transaction on Consumer Electronics, vol. 51, no. 1, pp , 25. [5] K. Hayashida, J. Ogawa, M. Nakayama, T. Nishiura, and Y. Yamashita, Multi-stage identification for abnormal/warning sounds detection based on maximum likelihood classification, ICA213, PaperID:1pSPb4, 213. [6] J. L. Rouas, J. Louradour, and S. Ambellouis, Audio events detection in public transport vehicle, IEEE Intelligent Transportation Systems Conference, pp , 26. [7] P. K. Atrey, N. C. Maddage, and M. S. Kankanhalli, Audio based event detection for multimedia surveillance, ICASSP 26, pp , 26. [8] S. Ntalampiras, I. Potamitis, and N. Fakotakis, An adaptive framework for acoustic monitoring of potential hazards, EURASIP Journal on Audio, Speech, and Music Processing, 29. [9] G. Valenzise, L. Gerosa, M. Tagliasacchi, F. Antonacci, and A. Sarti, Scream and gunshot detection and localization for audio-surveillance systems, IEEE Conference on Advanced Video and Signal Based Surveillance, pp , 27. [1] J. Pohjalainen, P. Alku, and T. Kinnunen, Shout detection in noise, ICASSP 211, pp , 211. [11] W. Huang, T. K. Chiew, H. Li, T. S. Kok, and J. Biswas, Scream detection for home applications, Industrial Electronics and Applications 21, pp , 21. [12],,,,,, 213, pp , 213. [13] J. Benesty, M. M. Sondhi, and Y. Huang, Springer handbook of processing, Springer, 28. [14] A. M. Noll, Cepstrum Pitch Determination, Journal of the Acoustical Society of America, vol. 41, no. 2, pp , [15] C Zhang and J.H.L Hansen Analysis and classification of mode: whispered through shouted, INTER- SPEECH 27, pp , 27. [16] HTK Software Toolkit, [17] Kaldi, [18] A. Varga and H.J.M. Steeneken, Assessment for automatic recognition: II. NOISEX-92: A database and an experiment to study the effect of additive noise on recognition systems, Speech Communication, vol. 12, no. 3, pp

1. HNS [1] HNS HNS HNS [2] HNS [3] [4] [5] HNS 16ch SNR [6] 1 16ch 1 3 SNR [4] [5] 2. 2 HNS API HNS CS27-HNS [1] (SOA) [7] API Web 2

1. HNS [1] HNS HNS HNS [2] HNS [3] [4] [5] HNS 16ch SNR [6] 1 16ch 1 3 SNR [4] [5] 2. 2 HNS API HNS CS27-HNS [1] (SOA) [7] API Web 2 THE INSTITUTE OF ELECTRONICS, INFORMATION AND COMMUNICATION ENGINEERS TECHNICAL REPORT OF IEICE. 657 8531 1 1 E-mail: {soda,matsubara}@ws.cs.kobe-u.ac.jp, {masa-n,shinsuke,shin,yosimoto}@cs.kobe-u.ac.jp,