Academic
Publications
Speech recognition in noisy environments using first-order vector Taylor series

Speech recognition in noisy environments using first-order vector Taylor series,10.1016/S0167-6393(97)00061-7,Speech Communication,Do Yeong Kim,Chong

Speech recognition in noisy environments using first-order vector Taylor series   (Citations: 48)
BibTex | RIS | RefWorks Download
Ž. In this paper, we generalize relations between clean and noisy speech signal using vector Taylor series VTS expansion Ž. for noise-robust speech recognition. We use it for both the noisy data compensation and hidden Markov model HMM parameter adaptation, and apply it for the cepstral domain directly, while Moreno used it to estimate the log-spectral parameters. Also, we develop a detailed procedure to estimate environmental variables in the cepstral domain using the Ž. Ž. expectation and maximization EM algorithms based on the maximum likelihood ML sense. To evaluate the developed method, we conduct speaker-independent isolated word and continuous speech recognition experiments. White Gaussian and driving car noises added to clean speech at various SNR are used as disturbing sources. Using only noise statistics obtained from three frames of silence and noisy speech to be recognized, we achieve significant performance improvement. Ž. Especially, HMM parameter adaptation with VTS is more effective than the parallel model combination PMC based on the log-normal assumption. q 1998 Elsevier Science B.V. All rights reserved. Resume ´´
Journal: Speech Communication , vol. 24, no. 1, pp. 39-49, 1998
Cumulative Annual
View Publication
The following links allow you to view full publications. These links are maintained by other sources not affiliated with Microsoft Academic Search.
    • ...In this case an auxiliary function is maximised using either second-order schemes, [5, 2, 6], or EM-based optimisation approaches [7, 4]. The latter solutions are closely related to generalisations of the Factor Analysis (FA) framework...
    • ...Eq. 4 has the form of a general FA-style model for which EM-based update formulae can be applied to estimate both clean or noise parameters [8, 7, 4, 3]. For example for noise parameter estimation, ˆ...
    • ...Σz. Note the convolutional noise, h, is completely described by its mean, ˆ μ h . It can again be estimated by maximising Eq. 7, yielding a similar expression to that obtained in [7]...

    F. Flegoet al. Factor analysis based VTS and JUD noise estimation and compensation

    • ...Well-known methods included in this last subcategory are cepstral mean normalization (CMN) [10], histogram equalization (HEQ) [11], [12], vector Taylor series (VTS) expansion [13], [14], and more recently a noise suppression algorithm based on the minimum mean square error (MMSE) criterion [15]...

    José A. Gonzálezet al. Efficient MMSE Estimation and Uncertainty Processing for Multienvironm...

    • ...Understandably, a simple linear approximation, namely the first-order vector Taylor series (VTS) approximation, has been tried in the past (e.g., [14], [16], and [17])...
    • ...In this appendix, we summarize how to derive, by extending the formulations in, e.g., [14] and [20], a procedure for the estimation of the parameters of explicit distortion model by maximizing the likelihood function defined on a given set of noisy observations in cepstral domain...

    Jun Duet al. A Feature Compensation Approach Using High-Order Vector Taylor Series ...

    • ...Examples of model-based methods are Vector Taylor Series for feature normalization (VTS) [6], Codeword Dependent Cepstral Normalization (CDCN) [7], and Spectral Subtraction (SS) [8]...

    Luis Bueraet al. Unsupervised Data-Driven Feature Vector Normalization With Acoustic Mo...

    • ...The first technique is based on the idea of “irrelevant variability normalization” (IVN) [17] and uses the VTS model adaptation in [13] as the basis for its approach...
    • ...Also, IVN is based on the VTS approach in [13], while NAT is based on the approach in [15]...
    • ...The goal of the traditional VTS model adaptation, e.g., [13]‐[15], is to adapt the parameters of the HMM trained using clean data to the environment conditions of a test utterance...
    • ...IVN is based on the VTS adaptation algorithm proposed in [13], while NAT uses VTS adaptation as performed in [14] and [15]...
    • ...In [13] (and thus in IVN), the log probability of the complete data used in the EM auxiliary function includes the observed noisy speech and hidden variables for clean speech, channel, noise and model component index (cf...
    • ...In [13] (and thus in IVN), the log probability of the complete data used in the EM auxiliary function includes the observed noisy speech and hidden variables for clean speech, channel, noise and model component index (cf. Equation (12) in [13])...
    • ...We note, however, that by using the auxiliary function in [13], IVNdoes havetheadvantage of closed-form update equations for both the means and variances, whereas an iterative approach is required for the variances in NAT...

    Ozlem Kalinliet al. Noise Adaptive Training for Robust Automatic Speech Recognition

Sort by: