The design of commercial spoken dialog systems is most commonly based on hand-crafting call flows. Voice interaction designers write prompts, predict caller responses, set speech recognition parameters, implement interaction strategies, all based on “best design practices”. Recently, we presented the mathematical framework “Contender” (similar to reinforcement learning) that allows for replacing manual decisions made during system design by data-driven soft decisions made at system run time optimizing the cumulative reward of an application. The current paper reports on the results of 26 Contenders implemented in commercial applications processing a total of about 15 million calls.

In commercial spoken dialog systems, call flows are built by call flow designers implementing a predefined business logic. While it may appear obvious from this logic how the call flow has to look like, i.e., which pieces of information have to be gathered from the caller or back-end systems and in which sequence, there are, in fact, strong arguments for automating call flow generation: 1) manual generation is time-consuming 2) manual generation is suboptimal and error-prone 3) automatic generation can react on dynamically changing business logic or external factors such as the distribution of callers and call reasons This paper presents a method for automatically deriving a call flow minimizing the average number of user turns given a business logic and a frequency distribution of call reasons. As an example, we applied the method to a call routing application whose manually built call flow is processing about 4 million calls per month and whose call reason distribution served to measure the impact of the automatic call flow generation.

We demonstrate three techniques (Escalator, Engager, and EverywhereContender) designed to optimize performance of commercial spoken dialog systems. These techniques have in common that they produce very small or no negative performance impact even during a potential experimental phase. This is because they can either be applied offline to data collected on a deployed system, or they can be incorporated conservatively such that only a low percentage of calls will get affected until the optimal strategy becomes apparent.

Contender (or what the academic community would refer to as a light version of reinforcement learning) is a simple technique to experiment with a number of competing paths in a (commercial) spoken dialog system. By randomly routing certain portions of traffic to individual paths and computing average rewards for each of the routes, the goal is to find out which one performs best. This paper is to do away with common uncertainties on how to set up contender weights, how much data needs to be accumulated to draw reliable conclusions, and how this all relates to the notion of statistical significance.

Statistical Spoken Language Understanding grammars (SSLUs) are often used only at the top recognition contexts of modern large-scale spoken dialog systems. We propose to use SSLUs at every recognition context in a dialog system, effectively replacing conventional, manually written grammars. Furthermore, we present a methodology of continuous improvement in which data are collected at every recognition context over an entire dialog system. These data are then used to automatically generate updated context-specific SSLUs at regular intervals and, in so doing, continually improve system performance over time. We have found that SSLUs significantly and consistently outperform even the most carefully designed rule-based grammars in a wide range of contexts in a corpus of over two million utterances collected for a complex call-routing and troubleshooting dialog system.

The localization of speech recognition for large-scale spoken dialog systems can be a tremendous exercise. Usually, all in- volved grammars have to be translated by a language expert, and new data has to be collected, transcribed, and annotated for statistical utterance classifiers resulting in a time-consuming and expensive undertaking. Often though, a vast number of transcribed and annotated utterances exists for the source lan- guage. In this paper, we propose to use such data and translate it into the target language using machine translation. The trans- lated utterances and their associated (original) annotations are then used to train statistical grammars for all contexts of the target system. As an example, we localize an English spoken dialog system for Internet troubleshooting to Spanish by trans- lating more than 4 million source utterances without any human intervention. In an application of the localized system to more than 10,000 utterances collected on a similar Spanish Internet troubleshooting system, we show that the overall accuracy was only 5.7% worse than that of the English source system. Index Terms: spoken dialog systems, machine translation, lo- calization

In this paper we introduce a subjective metric for evaluating the performance of spoken dialog systems, caller experience (CE). CE is a useful metric for tracking the overall performance of a system in deployment, as well as for isolating individual problematic calls in which the system underperforms. The proposed CE metric differs from most performance evaluation metrics proposed in the past in that it is a) a subjective, qualitative rating of the call, and b) provided by expert, external listeners, not the callers themselves. The results of an experiment in which a set of human experts listened to the same calls three times are presented. The fact that these results show a high level of agreement among different listeners, despite the subjective nature of the task, demonstrates the validity of using CE as a standard metric. Finally, an automated rating system using objective measures is shown to perform at the same high level as the humans. This is an important advance, since it provides a way to reduce the human labor costs associated with producing a reliable CE.

The annotation of hundreds of thousands of utterances for the training of statistical utterance classifiers requires a careful quality assurance procedure to make the data consistent and reliable. In this paper, we present five methods to analyze different aspects of annotated data to ensure their Completeness, Consistency, Correlation, Congruence and to avoid Confusion-collectively referred to as C5.

This paper compares six algorithms for call classification in the framework of a dialog system for automated troubleshooting. The comparison is carried out on large datasets, each consisting of over 100,000 utterances from two domains: television (TV) and Internet (INT). In spite of the high number of classes (79 for TV and 58 for INT), the best classifier (maximum entropy on word bigrams) achieved more than 77% classification accuracy on the TV dataset and 81% on the INT dataset.

In this paper we describe an unsupervised approach for the automated categorisation of utterances into predefined categ ories of symptoms (or problems) within the framework of a technical support automated agent. The utterance classification is pe rformed based on an iterative K-means clustering method. In order to improve the lower accuracy typical of unsupervised algorithms, we have analysed two different enhancements of the classification algorithm. The first method exploits the affin ity among words by automatically extracting classes of semantically equivalent terms. The second approach consists of a disambiguation technique based on a new criterion to estimate the relevance of terms for the classification. An analysis of the results of an experimental evaluation performed on a corpus of 34848 utterances concludes the paper. Index Terms: automated agents, automated categorisation, unsupervised learning, related words, term relevance.

In this paper we address the problem of optimizing the decisions taken by a spoken dialog system based on a given business model. Although the model can be applied to other types of decisions, we refer to the choice on whether to continue a session or to escalate it to a human agent. We consider a family of business models based on costs and savings that are dependent on the decision taken and the automation outcome of each session. Based on a corpus of more than 61,000 dialogs, we show that an optimal classifier can be selected to minimize the cost for a given business model.

Abstract ~ This paper describes a prototype of a conversational system that was implemented on the Ford Model U Concept Vehicle and first shown at the 2003 North American International Auto Show in Detroit. The system, including a touch screen and a speech recognizer, is used for controlling several non-critical automobile operations, such as climate, entertainment, navigation, and telephone. The pro-totype implements a natural language spoken dia-log interface integrated with an intuitive graphical user interface, as opposed to the traditional, speech only, command-and-control interfaces deployed in some of the vehicles currently on the market.

This paper proposes the use of Belief Networks (BN) for mixed-initiative dialog modeling. The BN-based framework was previously used for natural language understanding, where BNs infer the informational goal of the user's query based on its parsed semantic concepts. We extended this framework with the technique of backward inference that can automatically detect missing or spurious concepts based on the inferred goal. This is, in turn, used to drive the mixed-initiative dialog model that prompts for missing concepts and clarifies for spurious concepts. Applicability is demonstrated for a simple foreign exchange domain, and our framework's mixed-initiative interactions were shown to be superior to the system-initiative and user-initiative interactions. We also investigate the scalability and portability of the BN-based framework to the more complex air travel (ATIS) domain. Backward inference detected an increased number of missing and spurious concepts, which led to redundancies in the dialog model. We experimented with several remedial measures that showed promise in reducing the redundancies. We also present a set of principles for hand-assigning "degrees of belief" to the BN to reduce the demand for massive training data when porting to a new domain. Experimentation with the ATIS data also gave promising results.

This paper reports on the comparison between various acoustic feature sets and classification algorithms for classifying spoken utterances based on the emotional state of the speaker. The data set used for the analysis comes from a corpus of human-machine dialogs obtained from a commercial application. Emotion recognition is posed as a pattern recognition problem. We used three different techniques - linear discriminant classifier (LDC), k-nearest neighborhood (k-NN) classifier, and support vector machine classifier (SVC) -for classifying utterances into 2 emotion classes: negative and non-negative. In this study, two feature sets were used; the base feature set obtained from the utterance-level statistics of the pitch and energy of the speech, and the feature set analyzed by principal component analysis (PCA). PCA showed a performance comparable to the base feature sets. Overall, the LDC achieved the best performance with error rates of 27.54% on female data and 25.46% on males with the base feature set. The SVC, however, showed a better performance in the problem of data sparsity.

We describe a spoken dialog application framework that combines the power and flexibility of server-side Java Servlets and Java Server Pages (JSPs) with the deployment portability, reliability and scalability of standard web (HTTP) servers and VoiceXML clients. Applications are developed by extending a framework of Java classes in order to define dialogs through lower level actions such as speech recognition, audio prompting, speech synthesis, and backend data access. The framework delegates session data management to servlets, embedding frame-based representations for the application’s global and session data. Dialog flow is controlled through general constructions such as loops, conditionals, scoped sub-dialogs, along with scoped command, error, and exception handling. Prompting and grammars are configured through simple JSP templates that generate the VoiceXML instructions for the server to return to the client. The framework is designed to be extensible, as demonstrated by the implementation of customizable backup and repeat commands integrated with session data, command handling and grammar scoping.

Describes the use of belief networks (BNs) for dynamic semantic modeling within the United Airlines' flight information service (FLIFO). Callers can speak naturally to obtain status information about all flights (including arrival and departure times) of United Airlines. We aim at enabling the application to utilize dynamic call information to improve speech recognition performance. Dynamic call information include the location of the caller, the time and date of the call, and the caller's dialog history. Dynamic semantic models can incorporate such additional information about the call in rescoring the N-best recognition hypotheses. Our experiments showed that this improved the recognition accuracy of flight number utterances from 84.95% to 86.80%.

In this paper we explore the efficiency of various ways of expressing the form and meaning of natural language utterances as context-free grammars. We concentrate on the top-down parsing strategy employed in SpeechWorks 6.5, a strategy common to many systems. As with other search-based parsers, the key to efficiency is to limit the uncertainty of the parser at any given stage by reducing non-determinism in the grammar. Here we study the effects of different expressions of the same grammar in terms of efficiency. We also describe a methodology for transforming a source grammar into a more efficient expression of the same forms and meanings.

The design and implementation of the AT&T Communicator mixed-initiative spoken dialog system is described.The Communicator project, sponsored by DARPA and launched in 1999, is a multi-year multi-site project on advanced spoken dialog systems research.The main focus of this paper is on issues related to the design of mixed-initiative systems.In addition to describing our architecture and implementation of the complex travel task, the paper reports on some preliminary evaluation results.

This paper proposes the use of Belief Networks (BN) for mixed-initiative dialog modeling. The BN-based framework was previously used for natural language understanding, where BNs infer the informational goal of the user's query based on its parsed semantic concepts. We extended this framework with the technique of backward inference that can automatically detect missing or spurious concepts based on the inferred goal. This is, in turn, used to drive the mixed-initiative dialog model that prompts for missing concepts and clarifies for spurious concepts. Applicability is demonstrated for a simple foreign exchange domain, and our framework's mixed-initiative interactions were shown to be superior to the system-initiative and user-initiative interactions. We also investigate the scalability and portability of the BN-based framework to the more complex air travel (ATIS) domain. Backward inference detected an increased number of missing and spurious concepts, which led to redundancies in the dialog model. We experimented with several remedial measures that showed promise in reducing the redundancies. We also present a set of principles for hand-assigning degrees of belief to the BN to reduce the demand for massive training data when porting to a new domain. Experimentation with the ATIS data also gave promising results.

This paper describes our progress towards building a telephony-based spoken dialogue system for workshop/conference services. A mixed-initiative dialogue system has been developed that is engineered to o er users natural interaction with the system, ease-of-use and robustness towards ambiguous requests and machine errors. A prototype system, known as W99, is described in this paper

We propose a quantitative model for dialog systems that can be used for learning the dialog strategy. We claim that the problem of dialog design can be formalized as an optimization problem with an objective function reflecting different dialog dimensions relevant for a given application. We also show that any dialog system can be formally described as a sequential decision process in terms of its state space, action set, and strategy. With additional assumptions about the state transition probabilities and cost assignment, a dialog system can be mapped to a stochastic model known as Markov decision process (MDP). A variety of data driven algorithms for finding the optimal strategy (i.e., the one that optimizes the criterion) is available within the MDP framework, based on reinforcement learning. For an effective use of the available training data we propose a combination of supervised and reinforcement learning: the supervised learning is used to estimate a model of the user, i.e., the MDP parameters that quantify the user's behavior. Then a reinforcement learning algorithm is used to estimate the optimal strategy while the system interacts with the simulated user. This approach is tested for learning the strategy in an air travel information system (ATIS) task. The experimental results we present in this paper show that it is indeed possible to find a simple criterion, a state space representation, and a simulated user parameterization in order to automatically learn a relatively complex dialog behavior, similar to one that was heuristically designed by several research groups.

We introduce a stochastic model for dialogue systems based on Markov decision process. Within this framework we show that the problem of dialogue strategy design can be stated as an optimization problem, and solved by a variety of methods, including the reinforcement learning approach. The advantages of this new paradigm include objective evaluation of dialogue systems and their automatic design and adaptation. We show some preliminary results on learning a dialogue strategy for an air travel information system.

Recent progress in the field of spoken natural language understanding expanded the scope of spoken language systems to include mixed initiative dialogue. Currently there are no agreed upon theoretical foundations for the design of such systems. In this work we propose a stochastic model of computer-human interactions. This model can be used for learning and adaptation of the dialogue strategy and for objective evaluation. 1. INTRODUCTION Man-machine interactions, from simple touch tone menus to more complex speech based systems are ubiquitous in today's world. In the natural language and AI research communities there are even more complex examples of dialogue systems that make use of natural ways of communication like speech and language [1, 2, 3, 6]. In this paper we show that such dialogue systems can be formally described in terms of their state space, action set and strategy. The state of a dialogue system represents all the knowledge the system has at a certain time about interna...

Traditional stochastic language models for speech recognition (i.e. n-grams) are deterministic, in the sense that there is one and only one derivation for each given sentence. Moreover a fixed temporal window is always assumed in the estimation of the traditional stochastic language models. This paper shows how non-determinism is introduced to effectively approximate a back-off n-gram language model through a finite state network formalism. It also shows that a new flexible and powerful network formalization can be obtained by releasing the assumption of a fixed history size. As a result, a class of automata for language modeling (variable n-gram stochastic automata) is obtained, for which we propose some methods for the estimation of the transition probabilities. VNSAs have been used in a spontaneous speech recognizer for the ATIS task. The accuracy on a standard test set is presented.

The conventional n-gram language model exploits dependencies between words and their fixed-length past. This letter presents a model that represents sentences as a concatenation of variable-length sequences of units and describes an algorithm for unsupervised estimation of the model parameters. The approach is illustrated for the segmentation of sequences of letters into subword-like units. It is evaluated as a language model on a corpus of transcribed spoken sentences. Multigrams can provide a significantly lower test set perplexity than n-gram models.>

In this paper we discuss learning paradigms for the problem of understanding spoken language. The basic idea consists in redefining the language understanding problem in terms of translation between a natural language and a formal language that represents the meaning of sentences. Within this framework, with the assumption that input and output sentences can be put into sequential correspondence, understanding can be seen as a problem of sequential transduction. In this case several techniques exist for learning the corresponding transducers, some of which can be properly stated in terms of Hidden Markov modeling (conceptual HMMs). If the sequential assumption does not hold, there are new algorithms that also seem able to solve the learning problem. This view of a language understanding system opens new perspectives in the field of automatic learning of language models.

Recently, much interest has been generated regarding speech recognition systems based on Hidden Markov Models (HMMs) and neural network (NN) hybrids. Such systems attempt to combine the best features of both models: the temporal structure of HMMs and the discriminative power of neural networks. In this work we establish one more relation between the HMM and the NN paradigms by introducing the time-warping network (TWN) that is a generalization of both an HMM-based recognizer and a backpropagation net. The basic element of such a network, a time- warping neuron, extends the operation of the formal neuron of a backpropagation network by warping the input pattern to match it optimally to its weights. We show that a single-layer network of TW neurons is equivalent to a Gaussian density HMM-based recognition system. This equivalent neural representation suggests ways to improve the discriminative power of this system by using backpropagation discriminative training, and/or by generalizing the structure of the recognizer to a multi-layer net. The performance of the proposed network was evaluated on a highly confusable, isolated word, multi-speaker recognition task. The results indicate that not only does the recognition performance improve, but the separation between classes is enhanced, allowing us to set up a rejection criterion to improve the confidence of the system.

The authors extend the dynamic time warping (DTW) algorithm, widely used in automatic speech recognition (ASR), to a dynamic plane warping (DPW) algorithm, for application in the field of optical character recognition (OCR) or similar applications. Although direct application of the optimality principle reduced the computational complexity somewhat, the DPW (or image alignment) problem is exponential in the dimensions of the image. It is shown that by applying constraints to the image alignment problem, e.g., limiting the class of possible distortions, one can reduce the computational complexity dramatically, and find the optimal solution to the constrained problem in linear time. A statistical model, the planar hidden Markov model (PHMM), describing statistical properties of images is proposed. The PHMM approach was evaluated using a set of isolated handwritten digits. An overall digit recognition accuracy of 95% was achieved. It is expected that the advantage of this approach will be even more significant for harder tasks, such cursive-writing recognition and spotting.>

An understanding system, designed for both speech and text input, has been implemented based on statistical representation of task specific semantic knowledge. The core of the system is the conceptual decoder, which extracts the words and their association to the conceptual structure of the task directly from the acoustic signal. The conceptual information, which is also used to clarify the English sentences, is encoded following a statistical paradigm. A template generator and an SQL (structured query language) translator process the sentence and produce SQL code for querying a relational database. Results of the system on the official DARPA test are given.>

Integration of language constraints into a large vocabulary speech recognition system often leads to prohibitive complexity. We propose to factor the constraints into two components. The first is characterized by a covering grammar which is small and easily integrated into existing speech recognizers. The recognized string is then decoded by means of an efficient language post-processor in which the full set of constraints is imposed to correct possible errors introduced by the speech recognizer.

The authors report recent improvements to an HMM (hidden, Markov model)-based, continuous speech recognition system. These advances, which include the incorporation of interword context-dependent units and position-dependent units and an improved feature analysis, lead to a recognition system which gives a 95% word accuracy and 75% sentence accuracy for speaker independent recognition of the 1000-word, DARPA resource management task using the standard word pair grammar (with a perplexity of about 60). With the improved acoustic modeling of subword units, the overall error rate reduction was over 42% compared with the performance results reported in the baseline system. The best results obtained so far using the word pair grammar gave 95.2% average word accuracy for the three DARPA evaluation sets.>

The authors provide a detailed description of all aspects of the implementation of a large-vocabulary speaker-independent, continuous speech recognizer used as a tool for the development of recognition algorithms based on hidden Markov models (HMMs) and Viterbi decoding. The complexity of HMM recognizers is greatly increased by the introduction of detailed context-dependent units for representing interword coarticulation. A vectorized representation of the data structures involved in the decoding process, along with compilation of the connection information among temporally consecutive words and an efficient implementation of the beam search pruning, has led to a speedup of the algorithm of about one order of magnitude. A guided search can be used during a tuning phase for obtaining a speedup of more than three times. An average recognition time of about 25 s per sentence, although far from real time, allows one to perform a series of training experiments and to tune the recognition system parameters in order to obtain high word accuracy on complex recognition tasks such as the DARPA resource management task.>

Acoustic modeling method of basic speech subword units are discussed to provide high word recognition accuracy. It is shown that for a basic set of 47 context-independent phone-like units, word accuracies on the order of 86-90% can be obtained for a 1000-word vocabulary, in a speaker-independent mode, for a grammar with a perplexity of 60, on independent test sets. When the basic set of units is increased to include context-dependent units, word recognition accuracies of from 91 to 93% can be achieved on the same test sets. Based on outside results and some of the present ones, it is possible to increase word recognition accuracies by about 2-3% using known modeling techniques.>

A continuous-speech recognition and understanding system for a thousand-word vocabulary has been designed and implemented. It is able to answer queries put to a geographical database in natural Italian language. A discussion is presented of the recognition component of the system. It can produce a word lattice that is then processed by a syntactic-semantic component. In addition, a linguistic decoder exploiting word-pair constraints has been investigated. Its results have been compared to those obtained by similar approaches reported in the literature. The system relies on word preselection through lexical access by means of broad phonetic classes and on hidden Markov modeling of subword units. The improvements to the basic approach are presented and system performance is given. Average word accuracy and correct sentence recognition obtained for speaker-dependent tests performed by two speakers pronouncing 214 sentences are 94.5% and 89.3%, respectively. The perplexity of the word-pair language model is 25.>

Several techniques for dealing with the variability of recognition units are reported. The segmental k-means technique has been used for estimating two-state hidden Markov models for each of an inventory of 46 phones. Multiple models of these phones have been generated using clustering techniques intended to model separately acoustically distinct phenomena associated with the same unit. It is shown that increasing the number of models per phone can significantly increase the performance of a speaker dependent continuous speech recognizer, especially if model weights for phones which reflect their contexts in words found in the test vocabulary can be obtained. The authors feel that model splitting could be even more important in speaker dependent recognition, where it can be used to reduce model variability associated with multiple speakers as well as variable contexts. The results suggest that adequate training data must be available to train multiple models; otherwise, performance degrades when the number of models increases.>

A continuous speech recognition and understanding system is presented that accepts queries about a restricted geographical domain, expressed in free but syntactically correct natural language, with a lexicon of the order of one thousand words. A lattice of word candidates hypothesized by the speaker dependent recognition level is the interface to an understanding module that performs the syntactic and semantic analysis. The recognition subsystem generates word hypotheses by exploiting hidden Markov models of sub-word units. Bottom-up constraints are also introduced to restrict the set of candidate words. The understanding module determines the most likely sequence of words and represents its meaning in a parse-tree suitable to access a database. It makes use of a modified caseframe analysis driven by the word hypotheses likelihood scores. The results of a set of experiments performed in 150 sentences collected from one speaker are given.>

A system for recognizing isolated utterances belonging to a very large vocabulary is presented that follows a two-pass strategy. The first step, hypothesization, consists in the selection of a subset of word candidates, starting from the segmentation of speech into six broad phonetic classes. This module is implemented through a dynamic programming algorithm working in a three-dimensional space. The search is performed on a tree representing a coarse description of the lexicon. The second step is the search for the best N candidates according to a maximum-likelihood criterion. Each word candidate is represented by a graph of subword hidden Markov models, and a tree structure of the whole word subset is built on line for an efficient implementation of the Viterbi algorithm. A comparison with a direct approach that does not use the hypothesization module shows that the two-pass approach has the same performance with an 80% reduction in computational complexity.>

Recently a two step strategy for large vocabulary isolated word recognition has been successfully experimented. The first step consists in the hypothesization of a reduced set of word candidates on the basis of broad bottom-up features, while the second one is the verification of the hypotheses using more detailed phonetic knowledge. This paper deals with its extension to continuous speech. A tight integration between the two steps rather than a hierarchical approach has been investigated. The hypothesization and the verification modules are implemented as processes running in parallel. Both processes represent lexical knowledge by a tree. Each node of the hypothesization tree is labeled by one of 6 broad phonetic classes. The nodes of the verification tree are, instead, the states of sub-word HMMs. The two processes cooperate to detect word hypotheses along the sentence.>

The main problems with HMMs of sub-word units are the large amount of training data and computer time needed for estimating the parameters of the models. In some applications it is not proposable that a new speaker utters many hundreds of words to train the system, hence the interest arises for a quick adaptation based on some tens of training utterances. Two bounds are given for comparison with the results of the speaker adaptation, namely the recognition rates of speaker dependent and cross speaker recognition. Speaker dependent recognition is achieved by training the HMMs with nearly 1000 words uttered by the same speaker used in the tests. Cross speaker recognition, that gives a lower bound to the performance, concerns experiments in which the models were trained by a speaker different from that who uttered the test sentences. An adaptation algorithm using Parzen estimation and interpolation of the emission densities between the new and the old speaker models was investigated. It is able to give satisfactory recognition rates adapting the HMMs on the basis of only 40 training words uttered by the new speaker.

In this paper the problem of lexical access to large vocabularies by means of a coarse phonetic description of words is addressed. A generate and test approach is used: first a set of word candidates is extracted from the lexicon by means of a broad phonetic description of the input utterance, then a more detailed stochastic model of each word in this set, based on sub-word phonetic units, is obtained, and the likelihood of the candidate words is estimated using the Viterbi algorithm. Results of the application of the method to a large vocabulary isolated word recognition task are given. The candidate lists produced in the generation phase include the correct word in 98 times out of 100, their average size is of the order of 50 items for a 1011 word lexicon, while they do not exceed 300 units for a 13748 word lexicon.

This paper describes the design of a phonetic unit set for recognition of continuous speech where each unit is represented by an Hidden Markov Model. Starting from a unit set definition like classical diphones, many variations were made in order to have an improvement in recognition performance and a reduction in storage requirements. The definition of this unit set is presented, along with experimental comparisons with classical diphones and with phoneme-like units. The performance was qualitatively evaluated using the segmentation of the training data-base provided from the Viterbi algorithm in a forced recognition task. Classical recognition experiments have also been carried out using different "difficult vocabularies" as test.

This paper describes a connected speech recognition system based on Markov models. The performance of this system was analyzed and compared with that of a system which uses prototypes of words instead of Markov models. Some preliminary results are reported with reference to the recognition of connected digits.

It is well known that the isolated word recognition strategy based on pattern matching gives good performance; however, in order to achieve efficient implementation it is necessary to develop techniques to reduce computational complexity and memory requirements, especially when the vocabulary size is not very small. In this work three different pattern compression techniques are compared on the basis of efficiency as well as recognition performance when applied to pattern matching by means of dynamic programming in a speaker dependent context.