In this work we present a complete set of techniques devoted to offer robustness against frame losses in distributed speech recognition over packet-switched networks. The proposed scheme is composed of tree techniques, two of them are applied at the sender and the last one in the recognizer itself. On one hand, a media-specific forward error correction (FEC) technique is used to allow the recovery of information within the bursts. On the other hand, a recognizer-based technique well known by its remarkable ability to reduce the effects of long consecutive frame losses during recognition, the weighted Viterbi algorithm (WVA), is used to handle the additional information introduced by FEC codes. Moreover, a double stream strategy whereby interleaving can be applied along with FEC codes without any delay increase, is also applied. The application of interleaving allows to reduce the perceived burst length at the receiver, further improving the recognition performance. As a result, the proposed scheme can provide an acceptable performance even under extremely adverse channel conditions.

This paper focuses on the application of the Ramsey class of interleavers to the problem of achieving robust distributed speech recognition in the presence of burst-like packet loss. Although block interleavers of minimal latency have been already proposed by other authors in order to cope with this problem, the Ramsey class offers a more powerful kind of interleavers. Taking into account the behavior of the error concealment techniques commonly used in these recognition systems, it is possible to design Ramsey interleavers which, in comparison with the block interleavers of minimal latency, can counteract longer bursts with the same latency and thus improve the robustness of the recognition system. The effectiveness of the proposed design criterion is verified by means of simulations and experimental comparisons

This paper shows a complete volcano monitoring system that has been developed on the basis of the seismicity observed during three summer Antarctic surveys at Deception Island Volcano (Antarctica). The system is based on the state of the art in hidden Markov modelling (HMM) techniques successfully applied to other scenarios. A database containing a representative set of different seismic events including volcano-tectonic earthquakes, long-period events, volcanic tremor and hybrid events recorded during the 1994-1995 and 1995-1996 seismic surveys was collected for training and testing. Simple left-to-right HMMs and multivariate Gaussian probability density functions (PDF) with diagonal covariance matrix were used. The feature vector consists of the log-energies of a filter-bank consisting of 16 triangular weighting functions uniformly spaced between 0 and 20 Hz plus the first and second order derivatives. The system is suitable to operate in real-time and its accuracy is close to 90%. When the system was tested with a different data set including mainly long-period events registered during several seismic swarms during the 2001-2002 field survey, more than 95% of the recognized events were correctly marked by the recognition system

This paper presents GAIALAB, a Weblab project for digital communications distributed learning developed at the University of Granada (Spain). This initiative has been funded by a University program to improve teaching quality. The project addresses problems of teaching and learning innovatively and effectively by enabling the user interaction with the simulation engines through a Web browser. Our Weblab project provides an interface with JAVA applets that have been developed and integrated in the system. The Weblab project enables simulating digital communication systems and analyzing the results obtained through the Web. Moreover, all the applets that are included in the system have been designed to allow the modification of the simulation parameters, thus enabling a better understanding of the algorithms. The system also permits evaluating the experimental work developed by the students

We propose a model-based VAD derived from the Vector Taylor Series (VTS) approach. A Gaussian mixture (trained with clean speech) is used in order to provide an appropriate decision rule for speech/non-speech detection. Additionally, VTS approach adapts the Gaussian mixture to noise conditions, yielding a stable perfor- mance for a wide range of SNRs. We have evaluated its ability for speech/non-speech detection and also its application for robust speech recognition. When compared to other VAD methods, the proposed VAD shows the best trade-off in speech/non-speech de- tection. When applied for Wiener Filtering and for frame drop- ping, the proposed VAD also provides the best recognition results. Index Terms: voice activity detection (VAD), vector Taylor series approach (VTS), Gaussian mixture, Wiener filtering.

This letter shows an effective statistical voice activity detection algorithm based on the integrated bispectrum, which is defined as a cross spectrum between the signal and its square and inherits the ability of higher order statistics to detect signals in noise with many other additional advantages: 1) its computation as a cross spectrum leads to significant computational savings, and 2) the variance of the estimator is of the same order as that of the power spectrum estimator. The decision rule is formulated in terms of an average likelihood ratio test (LRT) involving successive integrated bispectrum speech features. With these and other innovations, the proposed method reports significant improvements in speech/pause discrimination as well as in speech recognition over standardized techniques such as ITU-T G.729, ETSI AMR, and AFE VADs, and over recently published VADs

This paper presents a mixed recovery scheme for robust distributed speech recognition (DSR) implemented over a packet channel which suffers packet losses. The scheme combines media-specific forward error correction (FEC) and error concealment (EC). Media-specific FEC is applied at the client side, where FEC bits representing strongly quantized versions of the speech vectors are introduced. At the server side, the information provided by those FEC bits is used by the EC algorithm to improve the recognition performance. We investigate the adaptation of two different EC techniques, namely minimum mean square error (MMSE) estimation, which operates at the decoding stage, and weighted Viterbi recognition (WVR), where EC is applied at the recognition stage, in order to be used along with FEC. The experimental results show that a significant increase in recognition accuracy can be obtained with very little bandwidth increase, which may be null in practice, and a limited increase in latency, which in any case is not so critical for an application such as DSR

Network-based speech recognition (NSR) and distributed speech recognition (DSR) have been proposed as solutions to translate speech recognition technologies to mobile environments. NSR is the most straightforward solution since it does not require any modification in the mobile phone, however DSR offers higher robustness against codec compression and transmission channel degradation. This paper explores an alternative approach for remote speech recognition which combines the advantages of NSR and DSR. In this scheme, a standard speech codec is used for speech transmission but the recognition is performed from the received codec parameters. In particular, we focus on the effect of transmission channel errors, which can cause a more severe performance reduction on speech recognition than codec distortion. First, we show that an NSR solution can approach DSR through a reconstruction technique along with an adapted noise reduction technique originally proposed for acoustic noise. Then, these results are improved by working with recognition features directly extracted from the codec bitstream by means of parameter transcoding. Required modifications on current networks in order to access the bitstream are described. The network upgrading with the tandem free operation (TFO) protocol is an attractive solution. This upgrade not only offers an overall improvement on the end-to-end speech quality, but would also allow a recognition performance similar, and even higher in poor channel conditions, to that obtained by DSR when parameter transcoding along with the proposed mitigation techniques are applied

This paper proposes a new packet loss concealment technique based on the inclusion in each packet of a few FEC bits, representing data replicas, combined with a minimum mean square error estimation (MMSE). This technique is developed for an Aurora-2 distributed speech recognition system working over an IP network. In addition to the data representing the transmitted speech frames, each packet includes some FEC bits representing a strongly VQ-quantized version (replicas) of previous and subsequent frames. When a loss burst occurs, the lost frames can be reconstructed from the VQ replicas. In order to mitigate the degradation introduced by the coarse VQ quantization of the replicas, a model-based MMSE estimation is applied. The experimental results show that, under a strongly degraded channel, it is possible to obtain up to 83.31 % of word accuracy with only 4 FEC bits or 88.47 % with 8 FEC bits per packet, when the Aurora mitigation algorithm only obtains 76.98 %.

Currently, there are technology barriers inhibiting speech processing systems that work in extremely noisy conditions from meeting the demands of modern applications. This letter presents a new voice activity detector (VAD) for improving speech detection robustness in noisy environments and the performance of speech recognition systems. The algorithm defines an optimum likelihood ratio test (LRT) involving multiple and independent observations. The so-defined decision rule reports significant improvements in speech/nonspeech discrimination accuracy over existing VAD methods that are defined on a single observation and need empirically tuned hangover mechanisms. The algorithm has an inherent delay that, for several applications, including robust speech recognition, does not represent a serious implementation obstacle. An analysis of the overlap between the distributions of the decision variable shows the improved robustness of the proposed approach by means of a clear reduction of the classification error as the number of observations is increased. The proposed strategy is also compared to different VAD methods, including the G.729, AMR, and AFE standards, as well as recently reported algorithms showing a sustained advantage in speech/nonspeech detection accuracy and speech recognition performance.

An effective voice activity detection (VAD) algorithm is proposed for improving speech recognition performance in noisy environments. The approach is based on the determination of the speech/nonspeech divergence by means of specialized order statistics filters (OSFs) working on the subband log-energies. This algorithm differs from many others in the way the decision rule is formulated. Instead of making the decision based on the current frame, it uses OSFs on the subband log-energies which significantly reduces the error probability when discriminating speech from nonspeech in a noisy signal. Clear improvements in speech/nonspeech discrimination accuracy demonstrate the effectiveness of the proposed VAD. It is shown that an increase of the OSF order leads to a better separation of the speech and noise distributions, thus allowing a more effective discrimination and a tradeoff between complexity and performance. The algorithm also incorporates a noise reduction block working in tandem with the VAD and showed to further improve its accuracy. A previous noise reduction block also improves the accuracy in detecting speech and nonspeech. The experimental analysis carried out on the AURORA databases and tasks provides an extensive performance evaluation together with an exhaustive comparison to the standard VADs such as ITU G.729, GSM AMR, and ETSI AFE for distributed speech recognition (DSR), and other recently reported VADs.

This paper describes a method of compensating for nonlinear distortions in speech representation caused by noise. The method described here is based on the histogram equalization method often used in digital image processing. Histogram equalization is applied to each component of the feature vector in order to improve the robustness of speech recognition systems. The paper describes how the proposed method can be applied to robust speech recognition and it is compared with other compensation techniques. The recognition experiments, including results in the AURORA II framework, demonstrate the effectiveness of histogram equalization when it is applied either alone or in combination with other compensation techniques.

This work addresses the mitigation of channel errors by means of efficient minimum mean-square-error (MMSE) estimation. Although powerful model-based implementations have been recently proposed, the computational burden involved can make them impractical. We propose two new approaches that maintain a good level of performance with a low computational complexity. These approaches keep the simple structure and complexity of a raw MMSE estimation, although they enhance it with additional source a priori knowledge. The proposed techniques are built on a distributed speech recognition system. Different degrees of tradeoff between recognition performance and computational complexity are obtained.

Network-based speech recognition (NSR) using the conventional speech channel with the enhanced full rate (EFR) or the adaptive multi-rate (AMR) codec is a very attractive approach since no change to existing mobile phones is needed. However, NSR reveals a degrading performance due to both transmission channel errors and the speech encoding process in comparison with distributed speech recognition (DSR), where speech features are efficiently coded and transmitted on a data channel. We focus on the degradation of the speech features caused by channel errors in an NSR system and propose methods to improve the quality of these features. Applying these methods, it turns out that the performance of an NSR system based on EFR coding is comparable to that based on DSR.

Currently, there are technology barriers inhibiting speech processing systems working under extreme noisy conditions. The emerging applications of speech technology, especially in the fields of wireless communications, digital hearing aids or speech recognition, are some examples of such systems often requiring a noise reduction technique in combination with a precise voice activity detector (VAD). This paper presents a new VAD for improving speech detection robustness in noisy environments and the performance of speech recognition systems. The algorithm uses long-term information about the speech signal to formulate the decision rule and estimates the subband SNR using specialized order statistics filters (OSF). The proposed algorithm is compared to the most commonly used VAD in the field, in terms of speech/nonspeech discrimination and also in terms of recognition performance when the VAD is used in an automatic speech recognition (ASR) system. Experimental results demonstrate a sustained advantage over different VAD methods including standard VAD such as G.729 and AMR which are used as a reference, the VAD of the Advanced Front-End (AFE) for distributed speech recognition (DSR), and recently reported algorithms.

The paper mainly focusses on an improved voice activity detection algorithm employing long-term signal processing and maximum spectral component tracking. The benefits of this approach have been analyzed in a previous work (Ramirez, J. et al., Proc. EUROSPEECH 2003, p.3041-4, 2003) with clear improvements in speech/non-speech discriminability and speech recognition performance in noisy environments. Two clear aspects are now considered. The first one, which improves the performance of the VAD in low noise conditions, considers an adaptive length frame window to track the long-term spectral components. The second one reduces misclassification errors in highly noisy environments by using a noise reduction stage before the long-term spectral tracking. Experimental results show clear improvements over different VAD methods in speech/pause discrimination and speech recognition performance. Particularly, improvements in recognition rate were reported when the proposed VAD replaced the VADs of the ETSI advanced front-end (AFE) for distributed speech recognition (DSR).

Currently, new trends in wireless communications are demanding reliable human-machine interaction in real-life environments. However, there are obstacles inhibiting automatic speech recognition systems (ASR) working in noisy environments. The main difficulty is the degradation suffered by ASR systems due to a mismatch between training and test conditions. This paper shows an improved voice activity detector (VAD) combining noise reduction and subband divergence estimation for improving the reliability of speech recognizers operating in noisy environments. The algorithm formulates the decision rule by measuring the divergence between the subband spectral magnitude of speech and noise using the Kullback-Leibler (KL) distance on the denoised signal. Experiments demonstrate a sustained advantage over different VAD methods including standard VADs such as G.729 and AMR, which are used as a reference, recently reported algorithms, and the VADs of the advanced frontend (AFE) for distributed speech recognition (DSR). 1.

Noise compensation methods for speech recognition provide a cleaned version of the speech representation. Usually this cleaned version is the expected value of the speech parameters given the observed noisy speech and the noise statistic. A more realistic representation should include the probability distribution of the cleaned speech instead of its expected value in order to represent the uncertainty associated to the compensation process due to the variability of the noise process. Recently, the inclusion of the uncertainty in the recognition process has been studied. Some approaches represent the uncertainty in the HMM parameters values. Other approaches represent it in the feature space. This second approach offers a much simpler system implementation and lower computational cost. In this paper we have developed a noise compensation technique that incorporates the variance of the cleaned speech into the speech representation. The variance is estimated using a Wiener filter during the speech feature enhancement process. This way of including the uncertainty implies the modification of the decoding rule. Experimental results using AURORA 2 database demonstrate a sustained improvement of the performance in the recognition system (about 21% word error rate reduction) when uncertainty is considered in the decoding rule.

This letter shows an innovative voice activity detector (VAD) based on the Kullback-Leibler (KL) divergence measure. The algorithm is evaluated in the context of the recently approved ETSI standard for distributed speech recognition (DSR). The VAD uses long-term information of the noisy speech signal in order to define a more robust decision rule yielding high accuracy. The mel-scaled filter bank log-energies (FBE) are modeled by means of Gaussian distributions, and a symmetric KL divergence is used for the estimation of the distance between speech and noise distributions. The decision rule is formulated in terms of the average subband KL divergence that is compared to a noise-adaptable threshold. An exhaustive analysis using the AURORA databases is conducted in order to assess the performance of the proposed method and to compare it to existing standard VAD methods.

This letter presents a new segmental nonlinear feature normalization algorithm to improve the robustness of speech recognition systems against variations of the acoustic environment. An experimental study of the best delay-performance tradeoff is conducted within the AURORA-2 framework, and a comparison with two commonly used normalization algorithms is presented. Computationally efficient algorithms based on order statistics are also presented. One of them is based on linear interpolation between sampling quantiles, and the other one is based on a point estimation of the probability distribution. The reduction in the computational cost does not degrade the performance significantly.

A study of combined source and channel coding applied to LSP parameters in wideband speech coding is presented. The traditional approach to protect against channel errors is to increase the bit-rate for channel coding, decreasing the bit-rate of the source coding according the channel conditions. Joint source-channel coding is an alternative that provides a technique to mitigate channel errors without an increase of the bit-rate due to channel coding. This paper presents a study of channel optimized vector quantizer and channel optimized matrix quantizer applied to line spectral pairs (LSP) parameters in wideband speech coding. Gaussian and slow-fading Rayleigh channels are considered and GMSK (Gaussian minimum shift-keying) is used as the modulation technique. In addition, for comparison purposes, the performance of other schemes (split vector quantization, split matrix quantization and split multistage vector quantization) for quantizing the LSP parameters are evaluated.

In this paper, we develop a new mitigation technique for a distributed speech recognition system over IP. We have designed and tested several methods to improve the interpolation used in the Aurora DSR ETSI standard without any significant increase of computational cost at the decoder. These methods make use of the information contained in the data-source, because, in IP networks, unlike in cellular networks, no information is received during packet losses. When a packet loss occurs, the lost information can be reconstructed through estimations from the N nearest received packets. Due to the enormous amount of combinations from previous and next received speech vector sequences, we have developed a methodology that drastically reduces the amount of required estimations.

This paper shows an efficient voice activity detector (VAD) that is based on the estimation of the long-term spectral diver- gence (LTSD) between noise and speech periods. The proposed method decomposes the input signal into overlapped speech frames, uses a sliding window to compute the long-term spec- tral envelope and measures the speech/non-speech LTSD, thus yielding a high discriminating decision rule and minimizing the average number of decision errors. In order to increase non- speech detection accuracy, the decision threshold is adapted to the measured noise energy while a controlled hang-over is ac- tivated only when the observed signal-to-noise ratio (SNR) is low. An exhaustive analysis of the proposed VAD is carried out using the AURORA TIdigits and SpeechDat-Car (SDC) databases. The proposed VAD is compared to the most com- monly used ones in the field in terms of speech/non-speech detection and recognition performance. Experimental results demonstrate a sustained advantage over G.729, AMR and AFE VADs.

This paper is mainly focused on showing experimental results of a feature extraction algorithm that combines spectral noise reduction and nonlinear feature normalization. The successfulness of this approach has been shown in a previous work, and in this one, we present several improvements that result in a performance comparable to that of the recently approved AFE for DSR. Noise reduction is now based on a Wiener filter instead of spectral subtraction. The voice activity detection based on the full-band energy has been replaced with a new one using spectral information. Relative improvements of 24.81% and 17.50% over our previous system are obtained for AURORA 2 and 3 respectively. Results for AURORA 2 are not as good as those for the AFE, but for AURORA 3 a relative improvement of 5.27% is obtained.

Various aspects relating to the human production and perception of speech have gradually been incorporated into automatic speech recognition systems. Nevertheless, the set of speech prosodic features has not yet been used in an explicit way in the recognition process itself. This study presents an analysis of prosody's three most important parameters, namely energy, fundamental frequency and duration, together with a method for incorporating this information into automatic speech recognition. On the basis of a preliminary analysis, a design is proposed for a prosodic feature classifier in which these parameters are associated with orthographic accentuation. Prosodic-accentual features are incorporated in a hidden Markov model recognizer; their theoretical formulation and experimental setup are then presented. Several experiments were conducted to show how the method performs with a Spanish continuous-speech database. Using this approach to process other database subsets, we obtained a word recognition error reduction rate of 28.91%.

The VTS approach for noise reduction is based on a statistical for mulation. It provides the expected value of the clean speech given the noisy observations and statistical models for the clean speech and the additive noise. The compensated signal is only an approximation of the clean one and retains a residual mismatch. The main objective of this work is to characterize this residual noise and to propose techniques to reduce its unwanted effects. Two different approaches to this problem are presented in this paper. The first one is based on linear filtering the time sequences of compensated acoustic parameters; for this purpose we use LDA-based RASTA-like FIR filters. The second approach is based on canceling the distortion introduced into the probability distribution of acoustic parameters and uses the well-known technique of histogram equalization. Results reported on AURORA database show that the proposed methods increase the recognition performance.

The noise usually produces a non-linear distortion of the feature space considered for Automatic Speech Recognition. This distortion causes a mismatch between the training and recognition conditions which significantly degrades the performance of speech recognizers. In this contribution we analyze the effect of the additive noise over cepstral based representations and we compare several approaches to compensate this effect. We discuss the importance of the non-linearities introduced by the noise and we propose a method (based on the histogram equalization technique) specifically oriented to the compensation of the non-linear transformation caused by the additive noise. The proposed method has been evaluated using the AURORA-2 database and task. The recognition results show significant improvements with respect to other compensation methods reported in the bibliography and reveals the importance of the non-linear effects of the noise and the utility of the proposed method.

This paper presents a new method for testing dialogue systems using a variety of real-world conditions simulated in lab. The method is based on the use of an additional dialogue system, called simulator, designed to behave as users interacting with the dialogue system to test. The behavior of the simulator is decided from diverse scenarios that represent user goals. The simulator tries to achieve the goals in the scenarios during the interaction with the dialogue system. We applied the method to test in noise conditions a dialogue system under development in our lab, considering white and babble noise. The method allowed to find out errors in the recognizer and in the strategy the system uses to handle user confirmations. Using the method, we tested the behavior of the system in terms of task completion, taking into account a VTS noise compensation technique. The experiments show that, if the VTS technique is not used, the average task completion is 31,43% for the white noise and 29,64% for the babble noise. If the VTS technique is used, the average task completion increases to 57,35% for the white noise and to 54,44% for the babble noise.

Distributed Speech Recognition involves the development of techniques to mitigate the degradations that the transmission channel introduces in the speech features. This work proposes an HMM framework from which different mitigation techniques oriented to bursty channels can be derived. In particular, two MMSE-based and a new Viterbi-based mitigation procedures are derived under this framework. Several implementation issues such as the channel SNR estimation or the application of hard decision on the received signal vectors are dealt with. Also, different boundary conditions suitable for the speech recognition application are studied for the different mitigation procedures. The experimental results show that the HMM-based techniques can effectively mitigate channel errors, even in very poor channel conditions.

This work is mainly focused on showing experimental results using a combination of two methods for noise compensation which are shown to be complementary: classical spectral subtraction algorithm and histogram equalization. While spectral subtraction is focused on the reduction of the additive noise in the spectral domain, histogram equalization is applied in the cepstral domain to compensate the remaining non-linear effects associated to channel distortion and additive noise. The estimation of the noise spectrum for the spectral subtraction method relies on a new algorithm for speech / non-speech detection (SND) based on order statistics. This SND classification is also used for dropping long speech pauses. Results on Aurora 2 and Aurora 3 are reported.

This paper presents a novel technique that allows testing spoken dialogue systems by means of an automatic generation of conversations. The technique permits to easily test spoken dialogue systems under a variety of lab-simulated conditions, as it is easy to vary or change the utterance corpus used to check the performance of the system. The technique is based on the use of a module called user simulator whose purpose is to behave as real users when they interact with dialogue systems. The behaviour of the simulator is decided by means of diverse scenarios that represent the goals of the users. The simulator aim is to achieve the goals set in the scenarios during the interaction with the dialogue system. We have applied the technique to test a dialogue system developed in our lab. The test has been carried out considering different levels of white and babble noise as well as a VTS noise compensation technique. The results prove that the dialogue system performance is worse under the babble noise conditions. The VTS technique has been effective when dealing with noisy utterances and has lead to better experimental results, particularly for the white noise. The technique has permitted to detect problems in the dialogue strategies employed to handle confirmation turns and recognition errors, suggesting that these strategies must be improved.

Combined source and channel coding is a technique to mitigate channel errors without increasing the bit error rate. Channel optimized vector quantizer (COVQ) performs these objectives in the context of vector quantization. This paper presents a study of channel optimized matrix quantizer (COMQ) applied to quantize the line spectral pair (LSP) parameters as an extension of COVQ technique. Gaussian and slow-fading Rayleigh channels are considered and GMSK (Gaussian minimum shift-keying) is used as modulation technique. Several channel signal to noise ratio (CSNR) are considered to measure the performance of this system. In addition, for comparison purposes, the performance of other schemes for quantizing the LSP parameters are computed.

This work presents a quantization technique for LSP parameters which results in a low bit-rate transmission while providing protection against channel errors. As a generalization of the so called Channel Optimized Vector Quantization (COVQ), Channel Optimized Matrix Quantization (COMQ) can remove intraframe and interframe LSP redundancy with the target of protecting the information sent through a channel in the presence of noise. Split COMQ is used in order to reduce storage requirements and complexity. Results show that Split COMQ gives better performance under certain error conditions and a lower bit rate transmission in all channel conditions compared to the reference quantization techniques.

In this paper we present a study about time-frequency distribution of acoustic-phonetic information for the Spanish language. This is based on a large Spanish database automatically labeled, and we conclude that results are similar to those obtained for hand-labeled english databases. We use bidimensional LDA [1] to extract discriminant features in time-frequency domain (TF) that are more robust in noise than the standard ones based on MFCC and time derivatives. We show that TF domain and its corresponding transformed domain (CTM) are equivalent from the point of view of LDA analysis and use this fact to reduce the dimensionality of the problem. Finally, cascade unidimensional LDA (CLDA) is applied first in frequency and then in time. This gives better estimates of projection vectors and better recognition performance. The proposed techniques are evaluated in a connected digit recognition task. Utterances have been artificially corrupted with additive real noises.

A channel optimized vector quantizer (COVQ) is studied for the case of transmission over waveform channels. In this work, a number of modulation schemes with multidimensional signal constellations are considered, specifically, results on the binary signalling. M-ary phase-shift keying (MPSK) and M-ary quadrature amplitude modulation (MQAM) performance using COVQ with hard-decision decoding, is optimized for additive white Gaussian noise (AWGN) and flat-fading Rayleigh channel. In addition, when a flat-fading Rayleigh channel is assumed, diversity techniques are used and evaluated to improve the performance of the system.

We present a study of a channel optimized matrix quantizer (COMQ) applied to quantize LSP (line spectral pair) parameters in a CELP (code excited linear prediction) coder. A modification of the DoD FS-1016 standard (Campbell et al., 1989) is used for this purpose. A Gaussian channel is considered as the channel through which information is sent and BPSK (binary phase shift-keying) is used as modulation technique. To measure the performance of this coder several channel signal to noise ratios are considered. Also, the performance of the FS-1016 standard coder is computed for comparison purposes.

This article describes a comparative study of language models in which the evaluation protocol has been set by AUPELF-UREF . We especially pay attention on the comparison between two methods of clustering words which are necessary in the design of the corresponding language models. The first classification is done by following a linguistic and theoretical method and the second one is based on an optimization method. Both methods are evaluated through the Shannon game. The vocabulary used is 20 000 words, the training corpus is made of two years of Le Monde newspaper (42M of words) and the test corpus (400 000 words) is extracted from 6 years of Le Monde Diplomatique. First evaluations show an improvement of 13% of recognized words in the first five ranks and a decrease of 25% in perplexity.This article describes a comparative study of language models in which the evaluation protocol has been set by AUPELF-UREF . We especially pay attention on the comparison between two methods of clustering words which are necessary in the design of the corresponding language models. The first classification is done by following a linguistic and theoretical method and the second one is based on an optimization method. Both methods are evaluated through the Shannon game. The vocabulary used is 20 000 words, the training corpus is made of two years of Le Monde newspaper (42M of words) and the test corpus (400 000 words) is extracted from 6 years of Le Monde Diplomatique. First evaluations show an improvement of 13% of recognized words in the first five ranks and a decrease of 25% in perplexity.This article describes a comparative study of language models in which the evaluation protocol has been set by AUPELF-UREF . We especially pay attention on the comparison between two methods of clustering words which are necessary in the design of the corresponding language models. The first classification is done by following a linguistic and theoretical method and the second one is based on an optimization method. Both methods are evaluated through the Shannon game. The vocabulary used is 20 000 words, the training corpus is made of two years of Le Monde newspaper (42M of words) and the test corpus (400 000 words) is extracted from 6 years of Le Monde Diplomatique. First evaluations show an improvement of 13% of recognized words in the first five ranks and a decrease of 25% in perplexity.This article describes a comparative study of language models in which the evaluation protocol has been set by AUPELF-UREF . We especially pay attention on the comparison between two methods of clustering words which are necessary in the design of the corresponding language models. The first classification is done by following a linguistic and theoretical method and the second one is based on an optimization method. Both methods are evaluated through the Shannon game. The vocabulary used is 20 000 words, the training corpus is made of two years of Le Monde newspaper (42M of words) and the test corpus (400 000 words) is extracted from 6 years of Le Monde Diplomatique. First evaluations show an improvement of 13% of recognized words in the first five ranks and a decrease of 25% in perplexity.

A new parameter for estimating the difficulty of a continuous speech recognition task, called speech decoding difficulty, is presented. It is obtained from the language model defined for the recognition task and the phonetic similarity between the transcriptions of the words that make up the vocabulary used. Two variants of the proposed task difficulty measure are introduced: ideal speech decoding difficulty (ISDD), which is not influenced by practical considerations on the recognition system implemented, and a second, more realistic variant, called practical speech decoding difficulty (PSDD) to study the performance of a specific recognition system when confronting a given task.

In this work we propose a novel way of discriminating the words that are recognized by a speech recognition system as correctly or incorrectly detected words. The procedure consists of the extraction of a set of characteristics for each word. Utilizing these characteristics, we have built two classifiers: the first one is a vector quantizer, while the second one, though also a vector quantizer, was trained using adaptative technique learning. The results obtained show an improvement in the performance of the recognizer achieved by reducing the number of insertions with no significant reduction in the correctly detected words.

It is usually assumed that grammar probabilities and acoustic probabilities in a continuous speech recognition system have to be incorporated to the general score with different weights. This is an experimental fact and there is no generally accepted theoretical explanation. We propose an explanation to this fact, related to the way grammar scoring is incorporated in the searching procedure. Accordingly to this explanation, we perform a set of experiments to test our hypothesis. We are also proposing a new way of introducing grammar probabilities in a tree-based vocabulary search strategy, where systems are usually bound to use the worst strategy. To apply our ideas to unigrams is rather simple. For more complex language models like bigrams we have to implement a new procedure.

The algorithms for the reduction of the number of features without degrading the performance of pattern recognition systems play an important role in real applications. A new algorithm for feature selection is proposed. This algorithm is based on the discriminative feature extraction (DFE) technique and has been applied to speech recognition. The experimental results show that the recognition systems accept important reductions of the number of features without a degradation of the performance. For the representation used in our experiments, the recognition error-rate is not significantly increased when the number of components in the feature vector is reduced from 42 to 20.

We present a preliminary version of a voice dialogue system suitable to deal with client orders and questions in fast-food restaurants. The system consists of two main sub-systems, namely a dialogue sub-system and a voice interface. The dialogue sub-system is a natural language processing system that may be considered a rule-based expert system, whose behaviour is decided from a recorded dialogue corpus obtained at a real restaurant. In this paper we present a general description of both sub-systems, and focus on knowledge representation, grammar, and module structure of the dialogue sub-system. An introduction to the natural language generation mechanism used is introduced, and future work is mentioned. Finally some conclusions are shown.

This work presents the STACC, Sistema Telef onico Autom atico de Consulta de Calificaciones (Automatic Telephone System for Consulting Marks). This system has been developed at our laboratory during 1996 and implements a service through telephone line that allows the students to consult by speech their marks after the exams by means of a simple phone call. This experience provided us an interesting point of view about the problems of real applications of speech technology. In this work we describe the system and some statistics about the use of STACC by the students are presented.

Signal representation is crucial for designinga speechrecognizer. The feature extractor selects the information to be used by the classifier to perform the recognition. In noisy environments, the data vectors representing the speech signal are changed and the recognizer performance is degraded by two main facts: (1) the mismatch between the training and the recognition conditions and (2) the degradation of the signal to be recognized. In such a situation, the representation of the speech signal plays an important role. In this paper, we analyze the importance of the representation for speechrecognition in noise. We apply the Discriminative Feature Extraction (DFE) method to optimize the representation. The experiments presented in this work show that the DFE method, which has been successfully applied in clean environments, leads also to improvements of the speech recognizers in noise.

Research on multiple vector quantization (MVQ) has shown the suitability of such a technique for speech recognition. Basically, MVQ proposes the use of one separate VQ codebook for each recognition unit. Thus, a MVQHMM model is composed of a VQ codebook and a discrete HMM model. This technique allows the incorporation in the recognition dynamics of the input sequence information wasted by discrete HMM models in the VQ process. The use of distinct codebooks also allows one to train them in a discriminative manner. We propose a new VQ codebook design method for MVQ-based systems, obtained from a modified maximum mutual information estimation. This method provides meaningful error reductions and is performed independently from the estimation of the discrete HMM part of the MVQ model. The results show that the proposed discriminative design turns the MVQHMM technique into a powerful acoustic modeling tool in comparison with other classical methods such as discrete or semicontinuous HMMs.

A continuous speech recognition system has to explore the search space with the only information given by the probabilities provided by a set of hidden Markov models representing the different sounds of the language and the restrictions imposed by the language model. The objective of this work is to include additional information, related to the estimated location of the boundary between two consecutive phonemes. This information is provided by an artificial recurrent neural network. 1. INTRODUCTION Continuous-speech automatic recognition systems normally use Viterbi or Baum-Welch algorithms [1] to determine the sentence of the accepted language which has most likely been uttered. The number of sentences generated by the grammar is usually very large. Thus the exploration of the search space requires large amounts of memory and time. In spite of the considerable effort devoted to the design of new search techniques [2], the size of search space still represents a major obstacle to the...

Although the continuous HMM (CHMM) technique seems to be the most flexible and complete tool for speech modeling, it is not always used for the implementation of speech recognition systems due to several problems related to training and computational complexity. Besides, it is not clear the superiority of continuous models over other well-known types of HMMs, such as discrete (DHMM) or semicontinuous (SCHMM) models, or multiple vector quantization (MVQ) models, a new type of HMM modeling. The authors propose a new variant of HMM models, the SCMVQ, HMM models (semicontinuous multiple vector quantization HMM), that uses one VQ codebook per recognition unit and several quantization candidates, Formally, SCMVQ modeling is the closest one to CHMM, although requiring less computation than SCHMMs. Besides, the authors show that SCMVQs can obtain better recognition results than DHMMs, SCHMMs or MVQs.>

Proposes a transform trellis coded quantization (TTCQ) scheme suitable for low-delay speech coding. This scheme is based on a general transform domain formulation for small frame sizes and the trellis coded quantization technique previously proposed where Ungerboeck's amplitude modulation trellises and set partitioning ideas are used for source coding. Using the discrete cosine transform the authors apply this technique to the coding of 7 kHz wideband speech, a matter of interest due to ISDN based applications. A 32 kbps low-delay coder is developed, simulation results showing a very high speech quality comparable to that obtained by the G722 standard at 64 kbps.>

This paper presents a new framework developed to apply Alphanets to CSR. For this purpose, a modular system is proposed. This system is made up by three different modules: LVQ module, SLHMM module and DP module. The SLHMM module is an expansion of an Alphanet, and therefore, can be interpreted as a HMM. The system can be trained globally applying backpropagation techniques. The used pruning procedure is based upon recognized units instead of observations, which reduces the number of nodes needed to recognize a sentence, compared to HMM-based systems using the same parameters for the models in both systems. Besides, the training procedure re-adapts the weights according to the new architecture in a few iterations since the initial parameters can be estimated from a classical HMM CSR system.>