강인한 음성인식을 위한 히스토그램 등화에 의한 특징보상 및 환경적응
Feature Compensation and Environmental Model Adaptation Utilizing Histogram Equalization for Robust Speech Recognition
3
조회
0
다운로드
0
인용
0
참고문헌
키워드
초록
음성인식 기술의 성공적인 상용화를 위해서는 주변잡음이나 채널왜곡에 오염된 음성신호에 대한 인식성능 저하를 효과적으로 방지할 수 있는 강인한 음성인식 기술의 개발이 필수적이다. 현재 대부분의 강인한 음성인식 기술들은 신호영역에서의 음성개선, 음성인식을 위한 특징영역에서의 특징보상, 그리고 음성인식기의 모델영역에서의 모델적응에 속한다. 본 논문에서는 가장 효율적인 특징보상 기술의 하나인 히스토그램 등화 기법의 성능개선을 위한 클래스 히스토그램 등화를 제안하고 아울러 모델영역에서 환경적응을 위한 히스토그램 등화의 적용에 대해 소개한다. Aurora2 음성 데이터베이스에 대한 성능평가에서 제안된 클래스 히스토그램 등화에 의한 특징보상은 기본 음성인식 특징인 멜 켑스트럼에 비해 약 61.17%의 오인식률 감소를 나타내었고 기존의 히스토그램 등화에 의한 특징보상과 비교할 때 약 19.61%의 의미있는 개선을 얻었다. 또한, 환경적응을 위한 히스토그램 등화는 멜 켑스트럼에 비해 약 62.84%의 오인식률 감소를 나타내었다.
Abstract
For successful commercialization of speech recognition technologies, it is highly required to develop robust speech recognition techniques for effectively reducing the performance degradation of speech recognizers in noisy environments. Most robust speech recognition techniques can be categorized into three major approaches such as speech enhancement in signal space, feature compensation in feature space, and model adaptation in model space. In this paper, we propose the class histogram equalization technique to improve the histogram equalization method which is known as one of the most efficient feature compensation techniques. Additionally, we present a histogram equalization-based environmental model adaptation approach. In the performance evaluation on the Aurora2 database, the class histogram equalization-based feature compensation technique produces a relative error reduction of 61.17% over the baseline mel-cepstral feature and 19.61% over the original histogram equalization-based feature compensation approach. Finally, the histogram equalization-based environmental model adaptation technique provides a relative error reduction of 62.84% over the baseline feature.
참고문헌 (17)
- X. Huang, "Spoken Language Processing: A Guide to Theory, Algorithm, and System Development" (2001), Prentice Hall
- A. Agarwal, "Two-stage mel-warped Wiener filter for robust speech recognition", Proc. ASRU (1999)
- Y. Ephraim, "Speech enhancement using a minimum mean-square error short-time spectral amplitude estimator", IEEE Trans. Acoust. Speech Signal Processing, 32(6), pp.1109~1121 (1984)
- Y. Ephraim, "Speech enhancement using a minimum mean-square error log-spectral amplitude estimator", IEEE Trans. Acoust. Speech Signal Processing, 33(2), pp.443~445 (1985)
- A. E. Rosenberg, "Cepstral channel normalization techniques for HMM-based speaker verification", Proc. ICSLP, pp.1835~1838 (1994)
- O. Viikki, "Cepstral domain segmental feature vector normalization for noise robust speech recognition", Speech Communication, 25, pp.133~147 (1998)
- S. Dharanipragada, "A nonlinear unsupervised adaptation technique for speech recognition", Proc. ICSLP, pp.556~559 (2000)
- P. J. Moreno, "A vector Taylor series approach for environment-independent speech recognition", Proc. Int. Conf. Acoustics, Speech, Signal Processing, pp.733~736 (1996)
- C. J. Leggetter, "Maximum likelihood linear regression for speaker adaptation of continuous density hidden Markov models", Computer Speech and Language, 9, pp.171~185 (1995)
- J.L. Gauvain, "Bayesian learning of Gaussian mixture densities for hidden Markov models", Proc. of the DARPA Speech and Natural Language Workshop, pp.101~120 (1991)
- R. Kuhn, "Rapid speaker adaptation in eigenvoice space", IEEE Trans. Speech and Audio Proc, 8(6), pp.695~707 (2000)
- M.J. Gales, "Robust continuous speech recognition using parallel model combination", IEEE Trans. Speech and Audio Processing, 4(5), pp.352~359 (1996)
- N. S. Kim, "Statistical linear approximation for environment compensation", IEEE Signal Processing Letters, 5(1), pp.8~10 (1998)
- J. C. Segura, "Cepstral domain segmental nonlinear feature transformations for robust speech recognition", IEEE Signal Processing Letters, 11, pp.517~520 (2004)
- A＇. de la Torre, "Histogram equalization of speech representation for robust speech recognition", IEEE Trans. Speech and Audio Processing, 13, pp.355~366 (2005)
- Y. Suh, "Probabilistic class histogram equalization for robust speech recognition", IEEE Signal Processing Letters, 14(4), pp.287~290 (2007)
- H.-G. Hirsch, "The Aurora experimental framework for the performance evaluation of speech recognition systems under noisy conditions", Proc. Int. Conf. Spoken Language Processing, pp.16~20 (2000)