Abstract: A system for understanding entries, such as speech, develops a classifier by employing prior knowledge with which a given corpus of training entries is enlarged threefold. A rule is created for each of the labels employed in the classifyier, and the created rules are applied to the given corpus to create a corpus of attachments by appending a weight of ?p(x), or 1??p(x), to labels of entries that meet, or fail to meet, respectively, conditions of the labels' rules, and to also create a corpus of non-attachments by appending a weight of 1??p(x), or ?p(x), to labels of entries that meet, or fail to meet conditions of the labels' rules.

Abstract: A system for understanding entries, such as speech, develops a classifier by employing prior knowledge with which a given corpus of training entries is enlarged threefold. The prior knowledge is embodied in a rule, combined from separate rules created for each label outputted by the classifier, each of which includes a weight measure p(x). A first a set of created entries for increasing the corpus of training entries is created by attaching all labels to each entry of the original corpus of training entries, with a weight &eegr;p(x), or &eegr;(1−p(x)), in association with each label that meets, or fails to meet, the condition specified for the label, &eegr; being a preselected positive number. The second set of is created by not attaching any of the labels to each of the original corpus of training entries, with a weight of &eegr;(1−p(x)), or &eegr;p(x), in association with each label that meets, or fails to meet, the condition specified for the label.

Abstract: Recognizing a stream of speech received as speech vectors over a lossy communications link includes constructing for a speech recognizer a series of speech vectors from packets received over a lossy packetized transmission link, wherein some of the packets associated with each speech vector are lost or corrupted during transmission. Each constructed speech vector is multi-dimensional and includes associated features. Potentially corrupted features within the speech vector are indicated to the speech recognizer when present. Speech recognition is attempted at the speech recognizer on the speech vectors when corrupted features are present. This recognition may be based only on certain or valid features within each speech vector. Retransmission of a missing or corrupted packet is requested when corrupted values are indicated by the indicating step and when the attempted recognition step fails.

Abstract: A method of rapidly training an automatic speech recognizer as part of a spoken dialog system for an enterprise includes extracting information from enterprise emails, web site content, and/or speech or data records of interactions between customers and the enterprise. The method comprises extracting the relevant data to develop a domain-specific language model, generating an acoustic model from non-domain-specific data, combining the domain-specific language model with the non-domain-specific acoustic model to initially deploy the spoken dialog service, and adapting the language models as task-specific data becomes available.

Abstract: A system and method are disclosed that improve automatic speech recognition in a spoken dialog system. The method comprises partitioning speech recognizer output into self-contained clauses, identifying a dialog act in each of the self-contained clauses, qualifying dialog acts by identifying a current domain object and/or a current domain action, and determining whether further qualification is possible for the current domain object and/or current domain action. If further qualification is possible, then the method comprises identifying another domain action and/or another domain object associated with the current domain object and/or current domain action, reassigning the another domain action and/or another domain object as the current domain action and/or current domain object and then recursively qualifying the new current domain action and/or current object. This process continues until nothing is left to qualify.

Abstract: A method and apparatus for speech recognition using second order statistics and linear estimation of cepstral coefficients. In one embodiment, a speech input signal is received and cepstral features are extracted. An answer is generated using the extracted cepstral features and a fixed signal independent diagonal matrix as the covariance matrix for the cepstral components of the speech input signal and, for example, a hidden Markov model. In another embodiment, a noisy speech input signal is received and a cepstral vector representing a clean speech input signal is generated based on the noisy speech input signal and an explicit linear minimum mean square error cepstral estimator.

Abstract: A multiple confidence measures subsystem of an automated speech recognition system allows otherwise independent confidence measures to be integrated and used for both training and testing on a consistent basis. Speech to be recognized is input to a speech recognizer and a recognition verifier of the multiple confidence measures subsystem. The speech recognizer generates one or more confidence measures. The speech recognizer preferably generates a misclassification error (MCE) distance as one of the confidence measures. The recognized speech output by the speech recognizer is input to the recognition verifier, which outputs one or more confidence measures. The recognition verifier preferably outputs a misverification error (MVE) distance as one of the confidence measures. The confidence measures output by the speech recognizer and the recognition verifier are normalized and then input to an integrator.

Abstract: A speech recognition system which effectively recognizes unknown speech from multiple acoustic environments includes a set of secondary models, each associated with one or more particular acoustic environments, integrated with a base set of recognition models. The speech recognition system is trained by making a set of secondary models in a first stage of training, and integrating the set of secondary models with a base set of recognition models in a second stage of training.

Abstract: Hidden Markov models (HMMs) rely on high-dimensional feature vectors to summarize the short-time properties of speech correlations between features that can arise when the speech signal is non-stationary or corrupted by noise. These correlations are modeled using factor analysis, a statistical method for dimensionality reduction. Factor analysis is used to model acoustic correlation in automatic speech recognition by introducing a small number of parameters to model the covariance structure of a speech signal. The parameters are estimated by an Expectation Maximization (EM) technique that can be embedded in the training procedures for the HMMs, and then further adjusted using Minimum Classification Error (MCE) training, which demonstrates better discrimination and produces more accurate recognition models.

Abstract: Hierarchical signal bias removal (HSBR) signal conditioning uses a codebook constructed from the set of recognition models and is updated as the recognition models are modified during recognition model training. As a result, HSBR signal conditioning and recognition model training are based on the same set of recognition model parameters, which provides significant reduction in recognition error rate for the speech recognition system.

Abstract: Speech recognition processing is compensated for improving robustness of speech recognition in the presence of enhanced speech signals. The compensation overcomes the adverse effects that speech signal enhancement may have on speech recognition performance, where speech signal enhancement causes acoustical mismatches between recognition models trained using unenhanced speech signals and feature data extracted from enhanced speech signals. Compensation is provided at the front end of an automatic speech recognition system by combining linear predictive coding and mel-based cepstral parameter analysis for computing cepstral features of transmitted speech signals used for speech recognition processing by selectively weighting mel-filter banks when processing frequency domain representations of the enhanced speech signals.

Abstract: In a speech recognition system, a recognition processor receives an unknown utterance signal as input. The recognition processor in response to the unknown utterance signal input accesses a recognition database and scores the utterance signal against recognition models in the recognition database to classify the unknown utterance and to generate a hypothesis speech signal. A verification processor receives the hypothesis speech signal as input to be verified. The verification processor accesses a verification database to test the hypothesis speech signal against verification models reflecting a preselected type of training stored in the verification database. Based on the verification test, the verification processor generates a confidence measure signal. The confidence measure signal can be compared against a verification threshold to determine the accuracy of the recognition decision made by the recognition processor.

Abstract: A signal bias removal (SBR) method based on the maximum likelihood estimation of the bias for minimizing undesirable effects in speech recognition systems is described. The technique is readily applicable in various architectures including discrete (vector-quantization based), semicontinuous and continuous-density Hidden Markov Model (HMM) systems. For example, the SBR method can be integrated into a discrete density HMM and applied to telephone speech recognition where the contamination due to extraneous signal components is unknown. To enable real-time implementation, a sequential method for the estimation of the bias (SSBR) is disclosed.