Search

The crucial importance of science and technology and its accurate peer reviewed dissemination, has once again been demonstrated during the current pandemic. Thus the COVID-19 pandemic together with the inevitable energy transition required by climate change, lead us to consider the issue of scientific and technical communication, both for the written papers and proceedings that have largely moved online (but not always in open access), and the various types of seminars, workshops, and symposia that frequently involve air travel with substantial CO2 impact. Online meetings that have become recently very popular, as well as online repositories for publications, themselves have a significant CO2---as well as environmental---impact, due to the massive use of electricity by information and communication technologies (ICT) and of the environmentally unfriendly manufacturing processes and decommissioning of ICT equipment. Presented is a broad overview of these aspects, and some recommendations regarding the future organization of scientific and technical communication, including: (1) peer-reviewed journals and proceedings with online open access; (2) the importance of face to face seminars and symposia, together with online meetings, to maintain the serendipity and importance of direct human contact while reducing the need for air travel; (3) the peer evaluation of research and of academic and research staff and its dependence on publications and their qualitative---rather than excessively quantitative---evaluation, where the concept of impact should include the usefulness of research to education, industry and society; (4) and the crucial role of ICT in all these aspects and the questions raised by the sustainability of ICT itself., info:eu-repo/semantics/published

International audience; In this paper, we introduce two new confidence measures for large vocabulary speech recognition systems. The major feature of these measures is that they can be computed without waiting for the end of the audio stream. We proposed two kinds of confidence measures: frame-synchronous and local. The frame-synchronous ones can be computed as soon as a frame is processed by the recognition engine and are based on a likelihood ratio. The local measures estimate a local posterior probability in the vicinity of the word to analyze. We evaluated our confidence measures within the framework of the automatic transcription of French broadcast news with the EER criterion. Our local measures achieved results very close to the best state-of-the-art measure (EER of 23% compared to 22.0%). We then conducted a preliminary experiment to assess the contribution of our confidence measure in improving the comprehension of an automatic transcription for the hearing impaired. We introduced several modalities to highlight words of low confidence in this transcription. We showed that these modalities used with our local confidence measure improved the comprehension of automatic transcription

International audience; This paper describes an extension of the n-gram language model: the similar n-gram language model. The estimation of the probability P(s) of a string s by the classical model of order n is computed using statistics of occurrences of the last n words of the string in the corpus, whereas the proposed model further uses all the strings s' for which the Levenshtein distance to s is smaller than a given threshold. The similarity between s and each string s' is estimated using co-occurrence statistics. The new P(s) is approximated by smoothing all the similar n-gram probabilities with a regression technique. A slight but statistically significant decrease in the word error rate is obtained on a state-of-the-art automatic speech recognition system when the similar n-gram language model is interpolated linearly with the n-gram model.

International audience; This paper presents several new confidence measures with the major advantage that they can be evaluated as soon as possible without having to wait for the recognition process to be completed. We have defined two kinds of confidence measures. The first one can be computed synchronously with the frame processed by the engine and the second one with a slight delay. Such measures are useful for driving the recognition process by modifying the likelihood score or for validating recognised words in on-the-fly applications such as keyword spotting task and on-line automatic speech transcription for deaf people. The EER evaluation on a French broadcast news corpus shows a performance close to the batch version of these measures (23.0% against 22.0% of EER) with only 0.84s of data before and after the word to be analysed.

Missing data recognition has been proposed to increase noiserobustness of automatic speech recognition. This strategy reliesontheuseofaspectrographicmaskthatgivesinformationaboutthetruecleanspeechenergyofacorruptedsignal. Thisinforma-tion is then used to refine the data process during the decodingstep. We propose in this work a new mask that provides moreinformation about the clean speech contribution than classicalmasks based on a Signal to Noise Ratio (SNR) thresholding.The proposed mask is described and compared to another miss-ing data approach based on SNR thresholding. Experimentalresults show a significant word error rate reduction induced bythe proposed approach. Moreover, the proposed mask outper-forms the ETSI advanced front-end on the HIWIRE corpus. IndexTerms : robustspeechrecognition,missingdata,boundedmarginalization 1. Introduction The presence of background noise typically causes mismatchesbetween training and testing conditions, which significantly de-gradetheperformanceofautomaticspeechrecognizers(ASRs).Over the last decades, many solutions to reduce the effect ofnoise have been proposed. Acoustic models can be adapted tonewnoisyconditions,theanalysisfront-endcanbemaderobustto noise, and noise reduction algorithms can be used as prepro-cessing stages.Although many of these methods have shown superior per-formance in noisy conditions compared to standard speechrecognition, noise robustness is still a challenging issue fornowadays speech recognizers, especially for non-sationarynoise.More recently, speech recognition with missing data hasbeen proposed. This technique relies on a clustering of spectralfeatures into two classes: time-frequency (T-F) units of a noisyspeechsignalthatcontainmorespeechenergythannoiseenergyare classified as reliable data, while T-F units containing morenoiseenergyareclassifiedasmissingdata. Hence, theresultingclustering produces a binary mask that is exploited in missingdata recognition techniques [1].

In this paper we present an automated approach for non-native speech recognition. We introduce a new phonetic confusion concept that associates sequences of native language (NL) phones to spoken language (SL) phones. Phonetic confusion rules are automatically extracted from a non-native speech database for a given NL and SL using both NL's and SL's ASR systems. These rules are used to modify the acoustic models (HMMs) of SL's ASR by adding acoustic models of NL's phones according to these rules. As pronunciation errors that non-native speakers produce depend on the writing of the words, we have also used graphemic constraints in the phonetic confusion extraction process. In the lexicon, the phones in words' pronunciations are linked to the corresponding graphemes (characters) of the word. In this way, the phonetic confusion is established between couples of (SL phones, graphemes) and sequences of NL phones. We evaluated our approach on French, Italian, Spanish and Greek non-native speech databases. The spoken language is English. The modified ASR system achieved significant improvements ranging from 20.3% to 43.2% (relative) in sentence error rate and from 26.6% to 50.0% in WER.

The agreement in gender and number is a critical problem in statistical language modeling. One of the main difficulties in speech recognition of French language is the presence of misrecognized words due to the bad agreement (in gender and number) between words. Statistical language models do not treat this phenomena directly. This paper focuses on how to handle the issue of this agreement. We introduce an original model called Features-Cache (FC) to estimate the gender and the number of the word to predict. It is a dynamic variable-length Features-Cache. The size of the cache is automatically determined in accordance to syntagm delimitors. The main advantage of this model is that there is no need to any syntactic parsing: it is used as any other statistical language model. Several models have been carried out and the best one achieves an improvement of approximatively 9 points in terms of perplexity. This model has been integrated in a speech recognition system based on JULIUS engine. Tests have been carried out on 280 sentences provided by AUPELF for the French automatic speech recognition evaluation campaign. This new model outperforms the baseline one, in terms of word error, by 3%.

This paper addresses the problem of spectrographic mask estimation in the context of missing data recognition. At the difference of other denoising methods, missing data recognition does not match the whole spectrum with the acoustic models, but rather considers that some time-frequency pixels are missing, i.e. corrupted by noise. Correctly estimating these ``masks'' is very important for missing data recognizers. We propose a new approach that exploits some a priori knowledge about these masks in typical noisy environments to address this difficult challenge. The proposed mask is then obtained by combining these noise dependent masks. The combination is led by an environmental ``sniffing'' module that estimates the probability of being in each typical noisy condition. This missing data mask estimation procedure has been integrated in a complete missing data recognizer using bounded marginalization. Our approach is evaluated on the Aurora2 database.

This paper presents a fully automated approach for the recognition of non-native speech based on acoustic model modification. For a native language (L1) and a spoken language (L2), pronunciation variants of the phones of L2 are automatically extracted from an existing non-native database as a confusion matrix with sequences of phones of L1. This is done using LTs and L2's ASR systems. This confusion concept deals with the problem of non existence of match between some L2 and L1 phones. The confusion matrix is then used to modify the acoustic models (HMMs) of L2 phones by integrating corresponding L1 phone models as alternative HMM paths. We introduce graphemic constraints in the confusion extraction process: the phonetic confusion is established for each couple of 'L2-phone' and the grapheme(s) corresponding to that phone. We claim that pronunciation errors may depend on the graphemes related to each phone. The modified ASR system achieved an improvement between 32% and 40% (relative, L1=French and L2=English) in WER on the French non-native database used for testing. The introduction of graphemic constraints in the phonetic confusion allowed further improvements

International audience; In this pap er, we describe our speech understanding system and we test it ontwo different applications. The proposed system is a task specific one and it concernespecially oral database consultation tasks. In this work, we consider that theautomatic speech understanding problem could be seen as an association problembetween two different languages. At the entry , the request expressed in naturallanguage and at the end, just before the interpretation stage, the same request isexpressed in term of concepts. A concept represents a given meaning, it is definedby a set of words sharing the same semantic properties. In this pap er, we propose anew Bayesian network based method to automatically extract the underlined concepts.We also propose and compare three approaches for the vector representationof words. We finish this pap er by a description of the post-processing step duringwhich we generate corresponding SQL queries to the pronounced sentences and weconnect our understanding system to a speech recognition engine. This step allowsus to validate our speech understanding approach by obtaining with the two treatedapplications the rates of 78% and 81% of well formed SQL requests.

This paper presents a fully automated approach for the recognition of non-native speech based on acoustic model modification. For a native language (L1) and a spoken language (L2), pronunciation variants of the phones of L2 are automatically extracted from an existing non-native database as a confusion matrix with sequences of phones of L1. This is done using L1's and L2's ASR systems. This confusion concept deals with the problem of non existence of match between some L2 and L1 phones. The confusion matrix is then used to modify the acoustic models (HMMs) of L2 phones by integrating corresponding L1 phone models as alternative HMM paths. In this way, no lexicon modification is carried. The modified ASR system achieved an improvement between 32% and 40% (relative, L1=French and L2=English) in WER on the French non-native database used for testing.

Colloque avec actes et comité de lecture. nationale.; National audience; In this paper, we compare the performance achieved by different word-spotting techniques based on hidden Markov models. We propose two methods to detect keywords, the first one uses a GMM (Gaussian Mixture Model) as a filler model to absorb the out-of-vocabulary words. The second is an alternative approach which does not attempt to model out-of-vocabulary words, instead, it uses buckled phonemes basedgrammar. Furthermore, it uses different reward functions to favourite the recognition of the keywords phonemes.

Article dans revue scientifique avec comité de lecture.; In natural language and especially in spontaneous speech, people often group words in order to constitute phrases which become usual expressions. This is due to phonological (to make the pronunciation easier), or to semantic reasons (to remember more easily a phrase by assigning a meaning to a block of words). Classical language models do not adequately take into account such phrases. A better approach consists in modeling some word sequences as if they were individual dictionary elements. Sequences are considered as additional entries of the vocabulary, on which language models are computed. In this paper, we present a method for automatically retrieving the most relevant phrases from a corpus of written sentences. The originality of our approach resides in the fact that the extracted phrases are obtained from a linguistically tagged corpus. Therefore, the obtained phrases are linguistically viable. To measure the contribution of classes in retrieving phrases, we have implemented the same algorithm without using classes. The class-based method outperformed by 11% the other method. Our approach uses information theoretic criteria which insure a high statistical consistency and make the decision of selecting a potential sequence optimal in accordance with the language perplexity. We propose several variants of language model with and without word sequences. Among them, we present a model in which the trigger pairs are linguistically more significant. We show that the use of sequences decrease the word error rate and improve the normalized perplexity. For instance, the best sequence model improves the perplexity by 16%, and the accuracy of our dictation system (MAUD) by approximately 14%. Experiments, in terms of perplexity and recognition rate, have been carried out on a vocabulary of 20000 words extracted from a corpus of 43 million words made up of two years of the French newspaper Le Monde. The acoustic model (HMM) is trained with the Bref80 corpus.

Colloque avec actes et comité de lecture. internationale.; International audience; The objective of this paper is to present a large study concerning the use of distant language models. In order to combine efficiently distant and classical models, an adaptation of the back-off principle is made. Also, we show the importance of each part of a history for the prediction. In fact, each sub-history is analyzed in order to estimate its importance in terms of prediction and then a weight is associated to each class of sub-histories. Therefore, the combined models take into account the features of each history's part and not the whole history as made in other works. The contribution of distant n-gram models in terms of perplexity is significant and improves the results by 12.8%. Making the linear combination depending on sub-histories achieves an improvement of $5.3\%$ in comparison to classical linear combination.

Colloque avec actes et comité de lecture. Internationale.; International audience; The automatic speech understanding problem could be considered as an association problem between two different languages. At the entry, the query expressed in oral or written natural language and at the end, just before the interpretation stage, the same request is expressed in term of concepts. One concept represents a given meaning, it is defined by a set of words sharing the same semantic properties. In this paper, we propose a new Bayesian network based method to automatically extract the underlined concepts. We also propose three different approaches for the vector representation of words. This representation allows the Bayesian network to build the adequate list of concepts for the considered application. This step is very important to obtain well built concepts. We finish this paper by a description of the post-processing step during which, we label our sentences and we generate the corresponding SQL queries. This step allows us to validate our automatic understanding approach and to obtain 92.5 of correct SQL queries on the test corpus.

Colloque avec actes et comité de lecture. internationale.; International audience; The automatic speech understanding problem could be considered as an association problem between two different languages. At the entry, the request expressed in natural language and at the end, just before the interpretation stage, the same request is expressed in term of concepts. A concept represents a given meaning, it is defined by a set of words sharing the same semantic properties. In this paper, we propose a new Bayesian network based method to automatically extract the underlined concepts. We also propose a new approach for the vector representation of words. We finish this paper by a description of the postprocessing step during which, we label our sentences and we generate the corresponding SQL queries. This step allows us to validate our speech understanding approach by obtaining good results. In fact, a rate of 92.5% of well formed SQL requests has been achieved on the test corpus.

Colloque avec actes et comité de lecture. internationale.; International audience; Support Vector Machines is a new and promising technique in statistical learning theory. Recently, this technique produced very interesting results in pattern recognition. In this paper, one of the first application of Support Vector Machines (SVM) technique for the problem of keyword spotting is presented. It classifies the correct and the incorrect keywords by using linear and Radial Basis Function kernels. This is a first work proposed to use SVM in keyword spotting, in order to improve recognition and rejection accuracy. The obtained results are very promising.

Colloque avec actes et comité de lecture. internationale.; International audience; Noise degrades the performance of Automatic Speech Recognition (ASR) systems working in real condition. The mismatch between the training and recognition conditions is considered the main factor involved in this degradation, and most methods for robust ASR are focussed on its minimization. In this work, we compare robust methods for ASR based on (a) the compensation of the noise effects and (b) the adaptation of the acoustic models to noise con-ditions. We propose a method for the adaptation of the acoustic models to the noise conditions based on a statistical formulation. In this method, each Gaussian is adapted to the noisy environment according to the estimated noise conditions. Recognition experiments have been carried out using speech acquired in real car environments. The results show the statistical formulation for adaptation provides an accurate method for robust ASR.

Colloque avec actes et comité de lecture. internationale.; International audience; In this paper, we present two methods for speech understanding : an artificial neural network and an information theory based method. For both methods we have to index input sentences using semantic classes (or concepts). In the first method, we perform supervised learning and we obtain very good indexing results. In the second one, we propose a new method based on mutual information statistical measure to retrieve concepts, and also to tag each sentence by its concepts. Both methods have been tested on a tourist information corpus. The information theory method yields better recall, whereas the neural network achieves a better precision. Better performance has been obtained by the neural network method (about 4%).

Colloque avec actes et comité de lecture. internationale.; International audience; This paper presents several statistical methods for topic identification on two kinds of textual data: newspaper articles and e-mails. Five methods are tested on these two corpora: topic unigrams, cache model, TFIDF classifier, topic perplexity, and weighted model. Our work aims to study these methods by confronting them to very different data. This study is very fruitful for our research. Statistical topic identification methods depend not only on a corpus, but also on its type. One of the methods achieves a topic identification of 80 % on a general newspaper corpus but does not exceed 30 % on e-mail corpus. Another method gives the best result on e-mails, but has not the same behavior on a newspaper corpus. We also show in this paper that almost all our methods achieve good results in retrieving the first two manually annotated labels.

Colloque avec actes et comité de lecture. internationale.; International audience; In this paper, we propose a new language model based on dependent word sequences organized in a multi-level hierarchy. We call this model MC$_{n}^{\nu}$, where $n$ is the maximum number of words in a sequence and $\nu$ is the maximum number of levels. The originality of this model is its capacity to take into account dependent variable-length sequences for very large vocabularies. In order to discover the variable-length sequences and to build the hierarchy, we use a set of $233$ syntactic classes extracted from the $8$ French elementary grammatical classes. The MC$_{n}^{\nu}$ model learns hierarchical word patterns and uses them to reevaluate and filter the n-best utterance hypotheses outputted by our speech recognizer MAUD. The model has been trained on a corpus of $43$ million words extracted from a French newspaper and uses a vocabulary of $20000$ words. Tests have been conducted on $300$ sentences. Results achieved $17\%$ decrease in perplexity compared to an interpolated class trigram model. Rescoring the original n-best hypotheses resulted in an improvement of $5\%$ in accuracy.

Colloque avec actes et comité de lecture. internationale.; International audience; In this paper we propose a new method to extract from a corpus the histories for which a given language model is better than another one. The decision is based on a measure stemmed from perplexity. This measure allows, for a given history, to compare two language models, and then to choose the best one for this history. Using this principle, and with a 20K vocabulary words, we combined two language models: a bigram and a distant bigram. The contribution of a distant bigram is significant and outperforms a bigram model by 7.5%. Moreover, the performance in Shannon game are improved. We show through this article that we proposed a cheaper framework in comparison to the maximum entropy principle, for combining language models. In addition, the selected histories for which a model is better than another one, have been collected and studied. Almost, all of them are beginnings of very frequently used French phrases. Finally, by using this principle, we achieve a better trigram model in terms of parameters and perplexity. This model is a combination of a bigram and a trigram based on a selected history.

Colloque avec actes et comité de lecture. internationale.; International audience; This paper focuses on language model adaptation, and more especially on topic identification (TID) for Automatic Speech Recognition (ASR). The structure of a set of topics is redefined by the introduction of a hierarchy. TID models may then make use of the semantic relationships between parent and son nodes of the topic-tree. The originality of the approach presented in this article lies in the allocation of a unique vocabulary to brother nodes, which rests on the use of two backing-off levels. In comparison with TID performance when using a non-hierarchical approach, results encourage us to carry on in this way.

Colloque avec actes et comité de lecture. internationale.; International audience; In this paper, an algorithm for continuous speech recognition systems based on the Multi-Band principle is proposed. This algorithm allows the bands to be asynchronous and has a practical complexity that is very close to the complexity of the classical Viterbi algorithm. The question of whether the bands should be constrained to be synchronous or not is discussed. We show that it is advantageous to let the bands asynchronous, as the increase of complexity, compared to the Viterbi algorithm, is low with our algorithm. Moreover, the accuracy must be at least as good as when the bands are synchronous, and, more importantly, different models than phones, can be used in the bands. In this paper, an algorithm for continuous speech recognition systems based on the Multi-Band principle is proposed. This algorithm allows the bands to be asynchronous and has a practical complexity that is very close to the complexity of the classical Viterbi algorithm. The question of whether the bands should be constrained to be synchronous or not is discussed. We show that it is advantageous to let the bands asynchronous, as the increase of complexity, compared to the Viterbi algorithm, is low with our algorithm. Moreover, the accuracy must be at least as good as when the bands are synchronous, and, more importantly, different models than phones, can be used in the bands.

Projet DIALOGUE; The coorelation between vectors in a sequence of analysis frames are supposed to be specific to phonetic units in acoustic-phonetic decoding of speech. We propose non-linear vector interpolation techniques to represent this correlation and to recognize phonemes. The interpolation is based on the decomposition of a frame sequence into two parts and on the construction of a function that interpolates one part using information from the second part. According to quantities to be interpolated, three families of interpolator models are developed. In a recognition system, each phonetic symbol is associated with a non-linear vector interpolator which is trained to give minimum interpolation error for that specific phoneme. Multi-layer feedforward neural networks are used to implement the non-linear vector interpolators. For a continuous speech phoneme spotting test using 16 LPCC-derived cepstrum coefficients as parametric vectors, the three categories of models gave compatible results. Vector-pair interpolator models yielded best recognition rate. Compared to a VQ-coded reference technique, this model gives close global recognition rate and significatly outperforms for plosive sounds.

X-TRA is a development tool for knowledge-based systems. It integrates a toolbox for truth maintenance based on both TMS and ATMS techniques. In this paper we present facilities provided in X-TRA for the truth maintenance task.

International audience; Automatic speech recognition and understanding (ASR) plays an important role in the framework of man-machine communication. Substantial industrial developments are at present in progress in this area. However, after 40 years or so of efforts several fundamental questions remain open. This paper is concerned with a comparative study of four different methods for multi-speaker word recognition: (i) clustering of acoustic templates, (ii) comparison with a finite state automaton, (iii) dynamic programming and vector quantization, (iv) stochastic Markov sources. In order to make things comparable, the four methods were tested with the same material made up of the ten digits (0 to 9) pronounced four times by 60 different speakers (30 males and 30 females). We will distinguish in our experiments between multi-speaker systems (capable of recognizing words pronounced by speakers that have been used during the training phase of the system) and speaker-independent systems (capable of recognizing words pronounced by speakers totally unknown to the system). Half of the corpus (15 male and 15 female) were used for training, and the remaining part for test.

International audience; Health information has encountered deep changes these last years under the influence of Information and Communication Technologies (ICT). Medical information that is controlled, produced and distributed by specialists (public and scientific institutions for example), has seen the rise of public information, partly produced by non-specialists. These evolutions have consequences on the access to this information, that has been enlarged and transformed, evolving from essentially professional uses for which the boundaries are clearly defined, to public uses in which personal discourses and specialized information are merged. This article raises two complementary facets of access to health information. Firstly, we present the professional practices, and those of patients whom participate in health forums.

International audience; In this article, we present an approach for non native automatic speech recognition (ASR). We propose two methods to adapt existing ASR systems to the non-native accents. The first method is based on the modification of acoustic models through integration of acoustic models from the mother tong. The phonemes of the target language are pronounced in a similar manner to the native language of speakers. We propose to combine the models of confused phonemes so that the ASR system could recognize both concurrent pronounciations. The second method we propose is a refinment of the pronounciation error detection through the introduction of graphemic constraints. Indeed, non native speakers may rely on the writing of words in their uttering. Thus, the pronounctiation errors might depend on the characters composing the words. The average error rate reduction that we observed is (22.5%) relative for the sentence error rate, and 34.5% (relative) in word error rate.