基于矢量泰勒级数的模型自适应算法
doi: 10.3724/SP.J.1146.2008.01768
Model Adaptation Algorithm Using Vector Taylor Series
-
摘要: 在实际环境中,由于测试环境与训练环境的不匹配,语音识别系统的性能会急剧恶化。模型自适应算法是减小环境失配影响的有效方法之一,它通过测试环境下的少量自适应数据,将HMM模型的参数变换到测试环境下。该文将矢量泰勒级数用于模型自适应,同时对HMM模型的均值向量和协方差矩阵进行变换,使其与实际环境相匹配。实验证明,该文算法优于MLLR算法和基于矢量泰勒级数的特征补偿算法,在低信噪比环境中性能提高尤为明显。Abstract: In actual environments the performance of speech recognition system may be degraded significantly because of the mismatch between the training and testing conditions. Model adaptation is an efficient approach that could reduce this mismatch, which adapts model parameters to new conditions by some adaptation data. In this paper, a new model adaptation using vector Taylor series is presented, which adapts the mean vector and covariance matrix of hidden Markov model. The experimental results show that the proposed algorithm is more effective than MLLR and the feature compensation algorithm based on vector Taylor series in various environments, especially in low signal-to-noise ratio environments.
-
[1] Moreno P J, Raj B, and Stern R M. A vector Taylor seriesapproach for environment- independent speech recognition[C].Proc. IEEE International Conference on Acoustics, Speech,and Signal Processing (ICASSP), Atlanta, Georgia, USA,7-10 May 1996: 733-736. [2] Moreno P J. Speech recognition in noisy environments[D].[Ph.D. dissertation], Carnegie Mellon University, 1996. [3] Sasou A, Asano F, and Nakamura S, et al.. HMM-basednoise-robust feature compensation[J].Speech Communication.2006, 48(9):1100-1111 [4] Kim W and Hansen J H L. Feature compensation in thecepstral domain employing model combination[J]. SpeechCommunication, 2009, 51(2): 83-96. [5] Gauvain J L and Lee C H. Maximum a posteriori estimationfor multivariate Gaussian mixture observations of Markovchains[J].IEEE Transactions on Speech and Audio Processing.1994, 2(2):291-298 [6] Leggetter C J and Woodland P C. Maximum likelihood linearregression for speaker adaptation of continuous densityhidden Markov models[J].Computer Speech and Language.1995, 9(2):171-185 [7] Gales M J F and Woodland P C. Mean and varianceadaptation within the MLLR framework[J].Computer Speechand Language.1996, 10(4):249-264 [8] Gales M J F and Young S J. Robust speech recognition inadditive and convolutional noise using parallel modelcombination[J].Computer Speech and Language.1995, 9(4):289-307 [9] Kim D and Yook D. Linear spectral transformation for robustspeech recognition using maximum mutual information[J].IEEE Signal Processing Letters.2007, 14(7):496-499 [10] Li J, Deng L, and Yu D, et al.. High-performance HMMadaptation with joint compensation of additive andconvolutive distortions via vector Taylor series[C]. Proc.IEEE Workshop on Automatic Speech Recognition andUnderstanding (ASRU), Kyoto, Japan, 9-13 December 2007:65-70.
计量
- 文章访问数: 3488
- HTML全文浏览量: 100
- PDF下载量: 1124
- 被引次数: 0