slide1 n.
Download
Skip this Video
Loading SlideShow in 5 Seconds..
outline PowerPoint Presentation
play fullscreen
1 / 15

outline

168 Views Download Presentation
Download Presentation

outline

- - - - - - - - - - - - - - - - - - - - - - - - - - - E N D - - - - - - - - - - - - - - - - - - - - - - - - - - -
Presentation Transcript

  1. COMBINED STATIC AND DYNAMIC VARIANCE ADAPTATION FOR EFFICIENT INTERCONNECTION OF SPEECH ENHANCEMENT PRE-PROCESSOR WITH SPEECH RECOGNIZER Marc Delcroix, Tomohiro Nakatani, Shinji WatanabeNTT Communication Science Laboratories, NTT Corporation A Maximum-Likelihood Approach to Stochastic Matching for Robust Speech Recognition (Ananth Sankar and Chin-Hui Lee Speech Research Department AT&T Bell Laboratories)

  2. outline Introduction The stochastic matching framework Estimation of Model-Space Transformation Dynamic variance compensation Proposed method for variance calculation Experiment Conclusion

  3. Introduction(3-1) • Conventionally there are two approaches for reducing the mismatch between the training and test conditions, namely model based approaches and feature based approaches. • The model adaptation relies on likelihood maximization, which assures a reduction in the mismatch. • Adaptive training is effective in removing static mismatch caused for example by speaker variations. However, it may not cope well with any mismatch arising for example from non-stationary noise or reverberation. • Feature based approaches consist of estimation clean speech feature using the observed speech. For example, speech enhancement methods can be used as a pre-processor to ASR.

  4. Introduction (3-2) • Many speech enhancement algorithms can efficiently reduce non-stationary noise. However, remaining noise or the excessive removal of noise may introduce distortions that prevent high recognition performance. • One kind of method uses the information on feature reliability to improve the performance of enhancement. • Focusing on reliable feature components during decoding • Dynamic variance compensation proposes increasing the model variance for unreliable feature components by adding the variance of enhanced feature. • So, there have been several proposals as regards estimating the variance of enhanced feature. • Usually depend on the speech enhancement pre-process • Lack generality. For example, feature variance is derived from a Gaussian mixture model of clean speech.

  5. Introduction (3-3) • The generality of the feature variance calculation could be increased by approximating it with the estimated observed noise. • However, the estimated variance may be far from the Oracle variance (i.e. the distance between clean and enhanced speech features). • Propose introducing a dynamic variance compensation scheme into a static adaptive training framework. • Design a parametric model for the feature variance that includes static and dynamic components. • Dynamic component is derived from the speech enhancement pre-processor output as the estimated observed noise. • can be performed for any pre-processor and assuring generality. • Parameters of variance model are optimized using an adaptive approach therefore may approach better Oracle feature variance.

  6. The stochastic matching framework (3-1) • The stochastic matching algorithm operates only on the given test utterance and the given set of speech models, and no additional training data is required for the estimation of the mismatch. • In speech recognition, the model are used to decode using the maximum a posteriori decoder • In the feature-space, if this distortion is invertible, then we may map back to the original speech with an inverse function , so that

  7. The stochastic matching framework (3-2) • In the model-based consider the transformation so that • One approach to decreasing the mismatch between and is to find the parameters or , and the word sequence that maximize the joint likelihood of and in • Thus in the feature-space, we need to find such that • Correspondingly, in the model-space, we need to find such that

  8. The stochastic matching framework (3-3) • To simplify expressions, we remove the dependence on , and write the maximum-likelihood estimation become • For this study, we assume that is a set of left to right continuous density subword HMMs, therefore the equation above can be written as • In general, it is not easy to estimate or directly. However, for some and , we can use the EM algorithm to iteratively improve on a current estimate and obtain a new estimate such that the likelihoods in equation above increase at each iteration.

  9. Estimation of Model-Space Transformation • Let the observation sequence be related to the original utterance and the distortion sequence by • Then, if the and are independent, we can write the probability density function of as • Let the statistical model of be given by ,which could be an HMM or a mixture Gaussian density.

  10. Dynamic variance compensation(2-1) • Recognition is usually achieved by finding a word sequence that maximize the function as: • Speech is modeled using a HMM with state density modeled by a Gaussian Mixture: • In practice, speech feature used for recognition may differ from clean speech features used for training, , because of noise, reverberation or distortions induced by speech enhancement pre-processing. In this paper, we focus on the latter case. • Model the mismatch, , between clean speech feature and enhanced speech feature as: where is modeled by a Gaussian as:

  11. Dynamic variance compensation(2-2) • The likelihood of a speech feature given a state , can be obtained by marginalizing the joint probability over mismatch as: where we assume the mismatch to be state independent, i.e.

  12. Proposed method for variance calculation(3-1) • In theory, feature variance should be computed as the squared difference between clean and pre-processed speech features. However, the clean speech features are unknown. • Assume that the feature variance is proportional to the estimated observed noise, i.e. the squared difference between observed noisy and pre-processed speech features. • Intuitively, this means that speech enhancement introduces more distortions when a great amount of noise is removed. One way to model feature variance is thus: where the is the Kronecker symbol, is the observed noisy speech, feature, is the enhanced feature and and are model parameters. • The model contains a dynamic variance part, , and a static bias, .

  13. For simplicity, we consider supervised adaptation, where the word sequence is known. The maximum likelihood estimation problem can be solved using the EM algorithm. We define an auxiliary function as where is a mismatch feature sequence. • The key point is to integrate dynamic variance concept into . • However, there is no closed form solution for the joint estimation of . Therefore, we consider the three following case, (i.e. static Variance Adaptation (SVA)), (dynamic Variance Adaptation (DVA)) and a combination of the two (SDVA). Proposed method for variance calculation(3-2)

  14. Proposed method for variance calculation(3-3) • SVA • DVA where