This paper investigates speech recognition with partial feature corruption, assuming unknown, time-varying noise characteristics. We extend our previous probabilistic union model from a conditional-probability formulation to a posterior-probability formulation. The new formulation allows the order of the model to be optimized for every single frame, and therefore greatly enhances the capability of the model for dealing with nonstationary noise corruption. Experiments have been conducted on two databases: TIDigits with noise corruption and Aurora 2, to demonstrate the improved robustness for the new model. Examples are presented showing that the new model can co-exist with existing noise-reduction techniques to provide improved noise robustness.

This paper introduces a new statistical approach, namely the probabilistic union model, for speech recognition involving partial, unknown frequency-band corruption. Partial frequency-band corruption accounts for the effect of a family of real-world noises. Previous methods based on the missing feature theory usually require the identity of the noisy bands. This identification can be difficult for unexpected noise with unknown, time-varying band characteristics. The new model combines the local frequency-band information based on the union of random events, to reduce the dependence of the model on information about the noise. This model partially accomplishes the target: offering robustness to partial frequency-band corruption, while requiring no information about the noise. This paper introduces the theory and implementation of the union model, and is focused on several important advances. These new developments include a new algorithm for automatic order selection, a generalization of the modeling principle to accommodate partial feature stream corruption, and a combination of the union model with conventional noise reduction techniques to deal with a mixture of stationary noise and unknown, nonstationary noise. For the evaluation, we used the TIDIGITS database for speaker-independent connected digit recognition. The utterances were corrupted by various types of additive noise, stationary or time-varying, assuming no knowledge about the noise characteristics. The results indicate that the new model offers significantly improved robustness in comparison to other models.

We present a new method of introducing domain knowledge into an n-gram language model. It is based on a combination of language models for individual word domains. Each word model is built from an individual corpus which is formed by extracting those subsets of the entire training corpus which contain that significant word. When testing, significant words are extracted from a cache and their models are combined with a global language model. Different methods of combining the models are described; one simple method based on combining frequencies rather than probabilities gives promising results and provides a relatively simple method of introducing domain information into an ngram language model. A 20% reduction in language model perplexity over the standard 3-gram approach is obtained which is similar to results obtained with other more complex domain models. The model also requires a small cache compared with other models requiring a cache.

A new statistical model is proposed for combining sub-band observations for robust speech recognition. This model characterizes the randomly corrupted sub-bands based on the union of random events. The new model has been incorporated into a hidden Markov model (HMM) and tested for recognizing a speaker-independent E-set, corrupted by various types of frequency-selective noise with unknown, time-varying statistics. The results show that the new model offers robustness to partial frequency corruption, requiring little or no knowledge about the noise characteristics.

Although the discrete cosine transform (DCT) is widely used for feature extraction in pattern recognition, it is shown that it converges slowly for most theoretically smooth functions. A modification of the DCT is described, based on a change of variable, which changes it to a new transform, called the discrete Chebyshev transform (DChT), which converges very rapidly for the same smooth functions. Although this rapid convergence is largely destroyed by the noise in real experimental data, the discrete Chebyshev transform is still generally better than the DCT when the sampling of the data can be selected at nonequidistant points. The improvement over the DCT gives a theoretical explanation for improved speech recognition obtained using Mel feature cepstral coefficients. These choose the sampling frequencies of a DCT to correspond to the human perception of pitch. It is shown that this sampling is similar to the sampling used in the discrete Chebyshev transform.

This paper provides a summary of our recent work on robust speech recognition based on a new statistical approach - the probabilistic union model. In particular, we considered speech recognition involving partial corruption in frequency bands, in time duration, and further in feature components. In all these situations, we assumed no prior knowledge about the corrupting noise, e.g. its band location, occurring time and statistical distribution. The new model characterizes these partial, unknown corruptions based on the union of random events. For the evaluation, we have conducted isolated-word recognition tasks by using both a speaker-independent E-set database and the TiDigits database, each being corrupted by various types of additive noise with unknown, time-varying statistics. The results indicate that the probabilistic union model offers robustness to partial corruption in speech utterances, requiring little or no knowledge about the noise characteristics.

Most current speech recognition systems are built upon a single type of model, e.g. an HMM or certain type of segment based model, and furthermore typically employs only one type of acoustic feature e.g. MFCCs and their variants. This entails that the system may not be robust should the modeling assumptions be violated. Recent research efforts have investigated the use of multi-scale/multi-band acoustic features for robust speech recognition. This paper described a multi-model approach as an alternative and complement to the multi-feature approaches. The multi-model approach seeks a combination of different types of acoustic models, thereby integrating the capabilities of each individual model for capturing discriminative information. An example system built upon the combination of the standard HMM technique with a segment-based modeling technique was implemented. Experiments for both isolated-word and continuous speech recognition have shown improved performances over each of the individual models considered in isolation.

A new method of using domains to improve a variable ngram language model is presented. Previous research on the use of domains using a 2-tier approach is extended through the addition of further hierarchical levels of domain abstraction. Our method is based upon the concept of domain association where semantically related domains exhibit similar n-gram distributions. Association allows us to cluster domains into a hierarchy of larger super-domains ultimately terminating when all domains have been included into one single superdomain. The resulting hierarchy has been used within a dynamic language model framework which tracks changes in text over time. Results with several corpora show improvements of 22% over traditional modelling techniques whilst providing a robust mechanism for handling topical changes.

This paper introduces a new statistical framework for constructing triphonic models from models of less context-dependency. This composition reduces the number of models to be estimated by higher than an order of magnitude and is therefore of great significance in relieving the data sparsity problem in triphone-based continuous speech recognition. The new framework is derived from Bayesian statistics, and represents an alternative to other triphone-by-composition techniques, particularly to the model-interpolation and quasitriphone approaches. The potential power of this new framework is explored by an implementation based on the hidden Markov modeling technique. It is shown that the new model structure includes the quasitriphone model as a special case, and leads to more efficient parameter estimation than the model-interpolation method. Phone recognition experiments show an increase in the accuracy over that obtained by comparable models.

A crucial issue in triphone based continuous speech recognition is the large number of models to be estimated against the limited availability of training data. This problem can be relieved by composing a triphone model from less context-dependent models. This paper introduces a new statistical framework, derived from the Bayesian principle, to perform such a composition. The potential power of this new framework is explored, both algorithmically and experimentally, by an implementation with hidden Markov modeling techniques. This implementation is applied to the recognition of the 39-phone set on the TIMIT database. The new model achieves 74.4% and 75.6% accuracy, respectively, on the core and complete test sets.

Two methods of improving conventional n-gram statistical language models are examined. The first involves using a new set of n-gram statistics that attempt to improve the ability of a system to identify phrases correctly. The second involves replacing the maximum likelihood unigram component with an optimised distribution. We test these approaches by incorporating them into weighted average [1] and deleted estimate [2] language models trained on a large newspaper corpus. The improvements lead to a reduction in perplexity of 4.5% and 4.9% respectively for these models.

An optimized hidden Markov model (HMM) with two kinds of inter-frame dependent observation structures, both built on the observation densities of a first-order dependent form, is presented to account for the statistical dependence between successive frames. In the first model, the dependence relation among the frames is determined optimally by maximizing the likelihood of the observations in both training and testing. In the second model, the dependence structure associated with each frame is described by a weighted sum of the conditional densities of the frame given individual previous frames. The segmental K-means and the forward-backward algorithms are implemented, respectively, for the estimation of the parameters of the two models. Experimental comparisons for an isolated word recognition task show that these models achieve better performance than both the standard continuous HMM and the bigram-constrained HMM.

A new hidden Markov model (HMM) with first-order dependent observation densities is presented to account for the statistical dependence between successive frames. In this model, the dependence relation among the frames is optimized to maximize the likelihood for both the training and testing data. Experimental comparisons with the standard continuous density HMM as well as the conditional-observation HMM for an isolated word recognition task show that a significant performance improvement is achieved for the new model.>

Disc access algorithms are discussed and compared in the context of on-line document retrieval. Emphasis is placed on algorithms which use the disc space efficiently and which minimise disc accesses, necessary features of viable on-line retrieval systems. One of the algorithms has been implemented in a pilot study of an interactive subject catalogue for part of the University Library at Belfast. Initial experience with this system is included in the paper.

An algorithm is described for the efficient and reliable evaluation of badly behaved definite integrals to a prescribed accuracy by concentrating the abscissas near the regions of greatest irregularity in the integrand. This is achieved by subdividing the interval of integration and by using a combination of the 7-point Clenshaw–Curtis quadrature and the 9-point Romberg quadrature in each subinterval. We argue that our algorithm will nearly minimise the number of function evaluations needed to evaluate a badly behaved integral.

An iterative method is described for the least-square curve fitting of a polynomial to a set of points in two dimensions when both the abscissas and ordinates are subject to error and when the weights of all the readings are known. The process converges, in general, to a polynomial giving the exact minimum of the ‘weighted’ perpendicular distances onto the curve. It is shown that in practice Deming's method gives a solution close to this optimum polynomial.

Several error estimates for the Clenshaw–Curtis quadrature formula are compared. Amongst these is one which is not unrealistically large, but which is easy to compute and reliable when certain conditions are satisfied. The form of this new error estimate helps explain the considerable accuracy of the Clenshaw–Curtis method when the integrand is well behaved; in this case the method is nearly as accurate as Gaussian quadratures. It is argued that the Clenshaw–Curtis method is a better method for evaluating such integrals than either Romberg's process or Gaussian quadratures.