In this work we present the Subsequence Similarity Language Model (S2-LM) which is a new approach to language modeling based on string similarity. As a language model, S2-LM generates scores based on the closest matching string given a very large corpus. In this paper we describe the properties and advantages of our approach and describe efficient methods to carry out its computation. We describe an n-best rescoring experiment intended to show that S2-LM can be adjusted to behave as an n-gram SLM model.

We present a new efficient algorithm for top-N match retrieval of sequential patterns. Our approach is based on an incremental approximation of the string edit distance using index information and a stack based search. Our approach produces hypotheses with average edit error of about 0.29 edits from the optimal SED result while using only about 5% of the CPU computation.

It is a common task in conversational applications to validate a claimed identity through utterance verification mechanisms; for example, by asking for a password or a challenge question. Sometimes, due to the private nature of the keyphrase, it is desirable to impose restrictions on how this keyphrase, or information that leads to it, is handled during the verification process. In this paper we describe a method where the utterance verification takes place in a concealed way, i.e., in a way that the keyphrase cannot be inferred or obtained directly by an intruder, from the decoding and application artifacts. We achieve this by a novel algorithm based on matrix operations in the verification network's nodes. If the traversed path corresponds to the keyphrase, the annotation converges to the dominant eigenvector of the occluded kernel matrix

Systems designed to extract time-critical information from large volumes of unstructured data must include the ability, both from an architectural and algorithmic point of view, to filter out unimportant data that might otherwise overwhelm the available resources. This paper presents an approach for data filtering to reduce computation in the context of a distributed speech processing architecture designed to detect or identify speakers. Here, filtering means either dropping and ignoring data or passing it on for further processing. The goal of the paper is to show that when the filter is designed to select and pass on a subset of the input data that best preserves the ability to recognize a specific desired speaker, or group of speakers, a large percentage of the data can be ignored while being able to preserve most of the accuracy

Dialog interaction in conversational applications is subordinated to the goal of completing a domain-specific task. In this paper we present a basic architecture, a knowledge representation system, and a planning algorithm for dialogue management that decouples the interaction process from the planning task. In our system, the interaction is driven by the planner. We use logic programming, automatic planning and problem solving algorithms for representing information states and performing interaction management in the dialogue system. Our approach leverages recent advances in formalisms, inference engines, planning and problem solving, and is particularly suitable when implementing negotiation-intensive conversational applications.

We describe a method for utterance classification for speech routing based on a discriminative training multinomial topic classifier. We propose the utilization of the n-norm a posteriori topic probability of the speech hypothesis as the objective function of a discriminative training step, and explore various ways to compute and maximize this function with respect to model parameters. To avoid obtaining negative probability estimates, we propose an alternative representation of the model parameterization. We utilize our approach in combination with a simple non-discriminative word detection algorithm based on Mutual Information and with a technique to identify the most salient phrases of the domain in order to alleviate the feature sparsity problem. We also explore the post-processing of classification results to further improve the classification performance via a decision tree model and a neural network. Overall, our discriminative trained multinomial system reduces the classification error rate up to 45% in an NLU task of financial transactions comparative to our baseline nondiscriminative multinomial system.

We describe a method to represent, manipulate, structure and aid in the design and development of NLU oriented parsers. Our method is based on the representation of the semantic parser domain into a single directed graph showing the parser's labels and their immediate inter-relationships as they exist in the annotated development corpora. Furthermore, we present methods developed around this representation illustrating how, a developer can visualize, manipulate, design and construct new applications simply by acting on the domain graph and sub-graphs. We also describe how the graph representation method can be utilized in the reduction of the complexity of the parser by identification and removal of nodes, edges and structures of the domain graph whose impact on attribute accuracy is small. We present the following examples of applications of our technique: extension of an existing air travel information domain to include car rental reservation by manipulating the corresponding graphs, structuring such graphs' vertices into 3-tiers, an example of a method for complex domain decomposition into simpler sub-graphs, and experiments on the reduction of a parser's complexity. Our technique can serve as a foundation of GUI toolkits for NLU development built around these concepts.

ABSTRACTIn this paper we describe some experiments on the Aurora 2 noisydigits database. The algorithms that we used can be broadly clas-sified into noise robustness techniques based on a linear-channelmodel of the acoustic environment such as CDCN [1] and its novelvariant termed Alignment-based CDCN( ACDCN , proposed here),and techniques which do not assume any particular knowledgeabout thestructure of the environment or noise conditions affectingthe speech signal such as discriminant feature space transforma-tions and speaker/channel adaptation. We present recognition ex-periments for both the clean training data and the multi-conditiontraining data scenarios.1. INTRODUCTIONIn this paper we describe the system and techniques for the Aurora2 noisy digits database and the results obtained. We developed twosets of acoustic models: the first set of models was trained on cleandata only and the second set was trained on the multi-conditiontraining data. For the system trained on clean data only, we appliedCDCN [1] and a novel variant of this technique called Alignment-based CDCN (

We describe procedures and experimental results using speech from diverse source languages to build an ASR system for a single target language. This work is intended to improve ASR in languages for which large amounts of training data are not available. We have developed both knowledge-based and automatic methods to map phonetic units from the source languages to the target language. We employed HMM adaptation techniques and discriminative model combination to combine acoustic models from the individual source languages for recognition of speech in the target language. Experiments are described in which Czech Broadcast News is transcribed using acoustic models trained from small amounts of Czech read speech augmented by English, Spanish, Russian, and Mandarin acoustic models.

In this paper we apply the Weighted Acoustic Modeling (WAM) technique to the recognition of speech coded by the full-rate GSM codec or the FS-1016 CELP codec employing various estimates of instantaneous distortion.In the WAM method, separate hidden Markov models are developed for regions of speech that exhibit low levels of codec-induced distortion and for regions with higher levels of such distortion.At recognition time, the contributions of these models are mixed together with a weighting that is determined by estimating the instantaneous distortion.In this paper instantaneous distortion was estimated from the instantaneous cepstral distortion, the long-term gain parameter of the codec, the long-term predictability of the reconstructed signal, and measurements of recoding sensitivity.We observe that the use of the long-term gain parameter produces results that are similar to those obtained by use of cepstral distortion (which can only be obtained if the original cepstra are transmitted along with the speech signal) for the GSM codec.Overall, the effect of the degradation in error rate introduced by coding can be reduced by up to 55% with these techniques for GSM coding, and by up to 38% for the CELP coding.

Speech coding affects speech recognition performance, with recognition accuracy deteriorating as the coded bit rate decreases.Virtually all systems that recognize coded speech reconstruct the speech waveform from the coded parameters, and then perform recognition (after possible noise and/or channel compensation) using conventional techniques.In this paper we compare the recognition accuracy of coded speech obtained by reconstructing the speech waveform with the speech recognition accuracy obtained when using cepstral features derived from the coding parameters.We focus our efforts on speech that has been coded using the 13-kbps full-rate GSM codec, a Regular Pulse Excited Long Term Prediction (RPE-LTP) codec.The GSM codec develops separate representations for the linear prediction (LPC) filter and the residual signal components of the coded speech.We measure the effects of quantization and coding on the accuracy with which these parameters are represented, and present two different methods for recombining them for speech recognition purposes.We observe that by selectively combining the cepstral streams representing the LPC parameters and the residual signal it is possible to obtain recognition accuracy directly from the coded parameters that equals or exceeds the recognition accuracy obtained from the reconstructed waveforms.