Recent works demonstrate that voice assistants do not perform equally well for everyone, but research on demographic robustness of speech technologies is still scarce. This is mainly due to the rarity of large datasets with controlled demographic tags. This paper introduces the Sonos Voice Control Bias Assessment Dataset, an open dataset composed of voice assistant requests for North American English in the music domain (1,038 speakers, 166 hours, 170k audio samples, with 9,040 unique labelled transcripts) with a controlled demographic diversity (gender, age, dialectal region and ethnicity). We also release a statistical demographic bias assessment methodology, at the univariate and multivariate levels, tailored to this specific use case and leveraging spoken language understanding metrics rather than transcription accuracy, which we believe is a better proxy for user experience. To demonstrate the capabilities of this dataset and statistical method to detect demographic bias, we consider a pair of state-of-the-art Automatic Speech Recognition and Spoken Language Understanding models. Results show statistically significant differences in performance across age, dialectal region and ethnicity. Multivariate tests are crucial to shed light on mixed effects between dialectal region, gender and age.
In recent years, there has been a significant increase in interest in developing Spoken Language Understanding (SLU) systems. SLU involves extracting a list of semantic information from the speech signal. A major issue for SLU systems is the lack of sufficient amount of bi-modal (audio and textual semantic annotation) training data. Existing SLU resources are mainly available in high-resource languages such as English, Mandarin and French. However, one of the current challenges concerning low-resourced languages is data collection and annotation. In this work, we present a new freely available corpus, named TARIC-SLU, composed of railway transport conversations in Tunisian dialect that is continuously annotated in dialogue acts and slots. We describe the semantic model of the dataset, the data and experiments conducted to build ASR-based and SLU-based baseline models. To facilitate its use, a complete recipe, including data preparation, training and evaluation scripts, has been built and will be integrated to SpeechBrain, a popular open-source conversational AI toolkit based on PyTorch.
Speech encoders pretrained through self-supervised learning (SSL) have demonstrated remarkable performance in various downstream tasks, including Spoken Language Understanding (SLU) and Automatic Speech Recognition (ASR). For instance, fine-tuning SSL models for such tasks has shown significant potential, leading to improvements in the SOTA performance across challenging datasets.In contrast to existing research, this paper contributes by comparing the effectiveness of SSL approaches in the context of (i) the low-resource Spoken Tunisian Arabic Dialect and (ii) its combination with a low-resource SLU and ASR scenario, where only a few semantic annotations are available for fine-tuning. We conducted experiments using many SSL speech encoders on the TARIC-SLU dataset. We used speech encoders that were pre-trained on either monolingual or multilingual speech data. Some of them have also been refined without in-domain nor Tunisian data through a multimodal supervised teacher-student learning. The study made in this paper yields numerous significant findings that we will discuss in the paper.
L’apprentissage auto-supervisé (SSL) a fait ses preuves pour le traitement automatique de la parole mais est généralement très consommateur de données, de mémoire et de ressources matérielles. L’approche BEST-RQ (BERT-based Speech pre-Training with Random-projection Quantizer) est une approche SSL performante en reconnaissance automatique de la parole (RAP), plus efficiente que wav2vec 2.0. L’article original de Google qui introduit BEST-RQ manque de détails, comme le nombre d’heures de GPU/TPU utilisées pour le pré-entraînement et il n’existe pas d’implémentation open-source facile à utiliser. De plus, BEST-RQ n’a pas été évalué sur d’autres tâches que la RAP et la traduction de la parole. Dans cet article, nous décrivons notre implémentation open-source de BEST-RQ et réalisons une première étude en le comparant à wav2vec 2.0 sur quatre tâches. Nous montrons que BERT-RQ peut atteindre des performances similaires à celles de wav2vec 2.0 tout en réduisant le temps d’apprentissage d’un facteur supérieur à deux.
La création de contenu journalistique peut être assistée par des outils technologiques comme la synthèse de parole. Cependant l’éditeur doit avoir la possibilité de contrôler la génération du contenu audio comme la prosodie, la prononciation ou le contenu linguistique. Dans ces travaux, un système de conversion de voix génère un signal de locuteur cible à partir d’une représentation temporelle de type Phonetic PosteriorGrams (PPGs) extraite d’un audio source. Les PPGs démêlent le contenu phonétique du contenu rythmique, et sont généralement considérés indépendants du locuteur. Cet article présente un système de conversion utilisant les PPGs, et son évaluation en qualité audio avec un test perceptif. Nous montrons également qu’un système de vérification du locuteur ne parvient pas à identifier le locuteur source après la conversion, même si le modèle a été entraîné sur des données synthétiques.
This paper reports on the shared tasks organized by the 20th IWSLT Conference. The shared tasks address 9 scientific challenges in spoken language translation: simultaneous and offline translation, automatic subtitling and dubbing, speech-to-speech translation, multilingual, dialect and low-resource speech translation, and formality control. The shared tasks attracted a total of 38 submissions by 31 teams. The growing interest towards spoken language translation is also witnessed by the constantly increasing number of shared task organizers and contributors to the overview paper, almost evenly distributed across industry and academia.
This paper describes the ON-TRAC consortium speech translation systems developed for IWSLT 2023 evaluation campaign. Overall, we participated in three speech translation tracks featured in the low-resource and dialect speech translation shared tasks, namely; i) spoken Tamasheq to written French, ii) spoken Pashto to written French, and iii) spoken Tunisian to written English. All our primary submissions are based on the end-to-end speech-to-text neural architecture using a pretrained SAMU-XLSR model as a speech encoder and a mbart model as a decoder. The SAMU-XLSR model is built from the XLS-R 128 in order to generate language agnostic sentence-level embeddings. This building is driven by the LaBSE model trained on multilingual text dataset. This architecture allows us to improve the input speech representations and achieve significant improvements compared to conventional end-to-end speech translation systems.
Though Dialogue State Tracking (DST) is a core component of spoken dialogue systems, recent work on this task mostly deals with chat corpora, disregarding the discrepancies between spoken and written language. In this paper, we propose OLISIA, a cascade system which integrates an Automatic Speech Recognition (ASR) model and a DST model. We introduce several adaptations in the ASR and DST modules to improve integration and robustness to spoken conversations. With these adaptations, our system ranked first in DSTC11 Track 3, a benchmark to evaluate spoken DST. We conduct an in-depth analysis of the results and find that normalizing the ASR outputs and adapting the DST inputs through data augmentation, along with increasing the pre-trained models size all play an important role in reducing the performance discrepancy between written and spoken conversations.
With the emergence of neural end-to-end approaches for spoken language understanding (SLU), a growing number of studies have been presented during these last three years on this topic. The major part of these works addresses the spoken language understanding domain through a simple task like speech intent detection. In this context, new benchmark datasets have also been produced and shared with the community related to this task. In this paper, we focus on the French MEDIA SLU dataset, distributed since 2005 and used as a benchmark dataset for a large number of research works. This dataset has been shown as being the most challenging one among those accessible to the research community. Distributed by ELRA, this corpus is free for academic research since 2019. Unfortunately, the MEDIA dataset is not really used beyond the French research community. To facilitate its use, a complete recipe, including data preparation, training and evaluation scripts, has been built and integrated to SpeechBrain, an already popular open-source and all-in-one conversational AI toolkit based on PyTorch. This recipe is presented in this paper. In addition, based on the feedback of some researchers who have worked on this dataset for several years, some corrections have been brought to the initial manual annotation: the new version of the data will also be integrated into the ELRA catalogue, as the original one. More, a significant amount of data collected during the construction of the MEDIA corpus in the 2000s was never used until now: we present the first results reached on this subset — also included in the MEDIA SpeechBrain recipe — , that will be used for now as the MEDIA test2. Last, we discuss evaluation issues.
In this paper we present two datasets for Tamasheq, a developing language mainly spoken in Mali and Niger. These two datasets were made available for the IWSLT 2022 low-resource speech translation track, and they consist of collections of radio recordings from daily broadcast news in Niger (Studio Kalangou) and Mali (Studio Tamani). We share (i) a massive amount of unlabeled audio data (671 hours) in five languages: French from Niger, Fulfulde, Hausa, Tamasheq and Zarma, and (ii) a smaller 17 hours parallel corpus of audio recordings in Tamasheq, with utterance-level translations in the French language. All this data is shared under the Creative Commons BY-NC-ND 3.0 license. We hope these resources will inspire the speech community to develop and benchmark models using the Tamasheq language.
Pretrained models through self-supervised learning have been recently introduced for both acoustic and language modeling. Applied to spoken language understanding tasks, these models have shown their great potential by improving the state-of-the-art performances on challenging benchmark datasets. In this paper, we present an error analysis reached by the use of such models on the French MEDIA benchmark dataset, known as being one of the most challenging benchmarks for the slot filling task among all the benchmarks accessible to the entire research community. One year ago, the state-of-art system reached a Concept Error Rate (CER) of 13.6% through the use of a end-to-end neural architecture. Some months later, a cascade approach based on the sequential use of a fine-tuned wav2vec2.0 model and a fine-tuned BERT model reaches a CER of 11.2%. This significant improvement raises questions about the type of errors that remain difficult to treat, but also about those that have been corrected using these models pre-trained through self-supervision learning on a large amount of data. This study brings some answers in order to better understand the limits of such models and open new perspectives to continue improving the performance.
The evaluation campaign of the 19th International Conference on Spoken Language Translation featured eight shared tasks: (i) Simultaneous speech translation, (ii) Offline speech translation, (iii) Speech to speech translation, (iv) Low-resource speech translation, (v) Multilingual speech translation, (vi) Dialect speech translation, (vii) Formality control for speech translation, (viii) Isometric speech translation. A total of 27 teams participated in at least one of the shared tasks. This paper details, for each shared task, the purpose of the task, the data that were released, the evaluation metrics that were applied, the submissions that were received and the results that were achieved.
This paper describes the ON-TRAC Consortium translation systems developed for two challenge tracks featured in the Evaluation Campaign of IWSLT 2022: low-resource and dialect speech translation. For the Tunisian Arabic-English dataset (low-resource and dialect tracks), we build an end-to-end model as our joint primary submission, and compare it against cascaded models that leverage a large fine-tuned wav2vec 2.0 model for ASR. Our results show that in our settings pipeline approaches are still very competitive, and that with the use of transfer learning, they can outperform end-to-end models for speech translation (ST). For the Tamasheq-French dataset (low-resource track) our primary submission leverages intermediate representations from a wav2vec 2.0 model trained on 234 hours of Tamasheq audio, while our contrastive model uses a French phonetic transcription of the Tamasheq audio as input in a Conformer speech translation architecture jointly trained on automatic speech recognition, ST and machine translation losses. Our results highlight that self-supervised models trained on smaller sets of target data are more effective to low-resource end-to-end ST fine-tuning, compared to large off-the-shelf models. Results also illustrate that even approximate phonetic transcriptions can improve ST scores.
This paper describes the ON-TRAC Consortium translation systems developed for two challenge tracks featured in the Evaluation Campaign of IWSLT 2021, low-resource speech translation and multilingual speech translation. The ON-TRAC Consortium is composed of researchers from three French academic laboratories and an industrial partner: LIA (Avignon Université), LIG (Université Grenoble Alpes), LIUM (Le Mans Université), and researchers from Airbus. A pipeline approach was explored for the low-resource speech translation task, using a hybrid HMM/TDNN automatic speech recognition system fed by wav2vec features, coupled to an NMT system. For the multilingual speech translation task, we investigated the us of a dual-decoder Transformer that jointly transcribes and translates an input speech. This model was trained in order to translate from multiple source languages to multiple target ones.
La reconnaissance des entités nommées (REN) à partir de la parole est traditionnellement effectuée par l’intermédiaire d’une chaîne de composants, exploitant un système de reconnaissance de la parole (RAP), puis un système de REN appliqué sur les transcriptions automatiques. Les dernières données disponibles pour la REN structurées à partir de la parole en français proviennent de la campagne d’évaluation ETAPE en 2012. Depuis la publication des résultats, des améliorations majeures ont été réalisées pour les systèmes de REN et de RAP. Notamment avec le développement des systèmes neuronaux. De plus, certains travaux montrent l’intérêt des approches de bout en bout pour la tâche de REN dans la parole. Nous proposons une étude des améliorations en RAP et REN dans le cadre d’une chaîne de composants, ainsi qu’une nouvelle approche en trois étapes. Nous explorons aussi les capacités d’une approche bout en bout pour la REN structurées. Enfin, nous comparons ces deux types d’approches à l’état de l’art de la campagne ETAPE. Nos résultats montrent l’intérêt de l’approche bout en bout, qui reste toutefois en deçà d’une chaîne de composants entièrement mise à jour.
Nous présentons un nouveau corpus, nommé AlloSat, composé de conversations en français extraites de centre d’appels, annotées de façon continue en frustration et satisfaction. Dans le contexte des centres d’appels, une conversation vise généralement à résoudre la demande de l’appelant. Ce corpus a été mis en place afin de développer de nouveaux systèmes capables de modéliser l’aspect continu de l’information sémantique et para-linguistique au niveau conversationnel. Nous nous concentrons sur le niveau para-linguistique, plus précisément sur l’expression des émotions. À notre connaissance, la plupart des corpus émotionnels contiennent des annotations en catégories discrètes ou dans des dimensions continues telles que l’activation ou la valence. Nous supposons que ces dimensions ne sont pas suffisamment liées à notre contexte. Pour résoudre ce problème, nous proposons un corpus permettant une connaissance en temps réel de l’axe frustration/satisfaction. AlloSat regroupe 303 conversations pour un total d’environ 37 heures d’audio, toutes enregistrées dans des environnements réels, collectées par Allo-Media (une société spécialisée dans l’analyse automatique d’appels). Les premières expériences de classification montrent que l’évolution de l’axe frustration/satisfaction peut être prédite automatiquement par conversation.
We present a new corpus, named AlloSat, composed of real-life call center conversations in French that is continuously annotated in frustration and satisfaction. This corpus has been set up to develop new systems able to model the continuous aspect of semantic and paralinguistic information at the conversation level. The present work focuses on the paralinguistic level, more precisely on the expression of emotions. In the call center industry, the conversation usually aims at solving the caller’s request. As far as we know, most emotional databases contain static annotations in discrete categories or in dimensions such as activation or valence. We hypothesize that these dimensions are not task-related enough. Moreover, static annotations do not enable to explore the temporal evolution of emotional states. To solve this issue, we propose a corpus with a rich annotation scheme enabling a real-time investigation of the axis frustration / satisfaction. AlloSat regroups 303 conversations with a total of approximately 37 hours of audio, all recorded in real-life environments collected by Allo-Media (an intelligent call tracking company). First regression experiments, with audio features, show that the evolution of frustration / satisfaction axis can be retrieved automatically at the conversation level.
This corpus is part of the PASTEL (Performing Automated Speech Transcription for Enhancing Learning) project aiming to explore the potential of synchronous speech transcription and application in specific teaching situations. It includes 10 hours of different lectures, manually transcribed and segmented. The main interest of this corpus lies in its multimodal aspect: in addition to speech, the courses were filmed and the written presentation supports (slides) are made available. The dataset may then serve researches in multiple fields, from speech and language to image and video processing. The dataset will be freely available to the research community. In this paper, we first describe in details the annotation protocol, including a detailed analysis of the manually labeled data. Then, we propose some possible use cases of the corpus with baseline results. The use cases concern scientific fields from both speech and text processing, with language model adaptation, thematic segmentation and transcription to slide alignment.
Named entity recognition (NER) from speech is usually made through a pipeline process that consists in (i) processing audio using an automatic speech recognition system (ASR) and (ii) applying a NER to the ASR outputs. The latest data available for named entity extraction from speech in French were produced during the ETAPE evaluation campaign in 2012. Since the publication of ETAPE’s campaign results, major improvements were done on NER and ASR systems, especially with the development of neural approaches for both of these components. In addition, recent studies have shown the capability of End-to-End (E2E) approach for NER / SLU tasks. In this paper, we propose a study of the improvements made in speech recognition and named entity recognition for pipeline approaches. For this type of systems, we propose an original 3-pass approach. We also explore the capability of an E2E system to do structured NER. Finally, we compare the performances of ETAPE’s systems (state-of-the-art systems in 2012) with the performances obtained using current technologies. The results show the interest of the E2E approach, which however remains below an updated pipeline approach.
In this paper, we propose several protocols to evaluate specific embeddings for Arabic sentiment analysis (SA) task. In fact, Arabic language is characterized by its agglutination and morphological richness contributing to great sparsity that could affect embedding quality. This work presents a study that compares embeddings based on words and lemmas in SA frame. We propose first to study the evolution of embedding models trained with different types of corpora (polar and non polar) and explore the variation between embeddings by observing the sentiment stability of neighbors in embedding spaces. Then, we evaluate embeddings with a neural architecture based on convolutional neural network (CNN). We make available our pre-trained embeddings to Arabic NLP research community with free to use. We provide also for free resources used to evaluate our embeddings. Experiments are done on the Large Arabic-Book Reviews (LABR) corpus in binary (positive/negative) classification frame. Our best result reaches 91.9%, that is higher than the best previous published one (91.5%).
Summarizing texts is not a straightforward task. Before even considering text summarization, one should determine what kind of summary is expected. How much should the information be compressed? Is it relevant to reformulate or should the summary stick to the original phrasing? State-of-the-art on automatic text summarization mostly revolves around news articles. We suggest that considering a wider variety of tasks would lead to an improvement in the field, in terms of generalization and robustness. We explore meeting summarization: generating reports from automatic transcriptions. Our work consists in segmenting and aligning transcriptions with respect to reports, to get a suitable dataset for neural summarization. Using a bootstrapping approach, we provide pre-alignments that are corrected by human annotators, making a validation set against which we evaluate automatic models. This consistently reduces annotators’ efforts by providing iteratively better pre-alignment and maximizes the corpus size by using annotations from our automatic alignment models. Evaluation is conducted on publicmeetings, a novel corpus of aligned public meetings. We report automatic alignment and summarization performances on this corpus and show that automatic alignment is relevant for data annotation since it leads to large improvement of almost +4 on all ROUGE scores on the summarization task.
This paper describes the ON-TRAC Consortium translation systems developed for two challenge tracks featured in the Evaluation Campaign of IWSLT 2020, offline speech translation and simultaneous speech translation. ON-TRAC Consortium is composed of researchers from three French academic laboratories: LIA (Avignon Université), LIG (Université Grenoble Alpes), and LIUM (Le Mans Université). Attention-based encoder-decoder models, trained end-to-end, were used for our submissions to the offline speech translation track. Our contributions focused on data augmentation and ensembling of multiple models. In the simultaneous speech translation track, we build on Transformer-based wait-k models for the text-to-text subtask. For speech-to-text simultaneous translation, we attach a wait-k MT system to a hybrid ASR system. We propose an algorithm to control the latency of the ASR+MT cascade and achieve a good latency-quality trade-off on both subtasks.
Dans cet article, nous présentons une approche de bout en bout d’extraction de concepts sémantiques de la parole. En particulier, nous mettons en avant l’apport d’une chaîne d’apprentissage successif pilotée par une stratégie de curriculum d’apprentissage. Dans la chaîne d’apprentissage mise en place, nous exploitons des données françaises annotées en entités nommées que nous supposons être des concepts plus génériques que les concepts sémantiques liés à une application informatique spécifique. Dans cette étude, il s’agit d’extraire des concepts sémantiques dans le cadre de la tâche MEDIA. Pour renforcer le système proposé, nous exploitons aussi des stratégies d’augmentation de données, un modèle de langage 5-gramme, ainsi qu’un mode étoile aidant le système à se concentrer sur les concepts et leurs valeurs lors de l’apprentissage. Les résultats montrent un intérêt à l’utilisation des données d’entités nommées, permettant un gain relatif allant jusqu’à 6,5 %.
Malgré les faiblesses connues de cette métrique, les performances de différents systèmes de reconnaissance automatique de la parole sont généralement comparées à l’aide du taux d’erreur sur les mots. Les transcriptions automatiques de ces systèmes sont de plus en plus exploitables et utilisées dans des systèmes complexes de traitement automatique du langage naturel, par exemple pour la traduction automatique, l’indexation, la recherche documentaire... Des études récentes ont proposé des métriques permettant de comparer la qualité des transcriptions automatiques de différents systèmes en fonction de la tâche visée. Dans cette étude nous souhaitons mesurer, qualitativement, l’apport de l’adaptation automatique des modèles de langage au domaine visé par un cours magistral. Les transcriptions du discours de l’enseignant peuvent servir de support à la navigation dans le document vidéo du cours magistral ou permettre l’enrichissement de son contenu pédagogique. C’est à-travers le prisme de ces deux tâches que nous évaluons l’apport de l’adaptation du modèle de langage. Les expériences ont été menées sur un corpus de cours magistraux et montrent combien le taux d’erreur sur les mots est une métrique insuffisante qui masque les apports effectifs de l’adaptation des modèles de langage.
Nous nous intéressons, dans cet article, à la tâche d’analyse d’opinions en arabe. Nous étudions la spécificité de la langue arabe pour la détection de polarité. Nous nous focalisons ici sur les caractéristiques d’agglutination et de richesse morphologique de cette langue. Nous avons particulièrement étudié différentes représentations d’unité lexicale : token, lemme et light stemme. Nous avons construit et testé des espaces continus de ces différentes représentations lexicales. Nous avons mesuré l’apport de tels types de representations vectorielles dans notre cadre spécifique. Les performances du réseau CNN montrent un gain significatif de 2% par rapport à l’état de l’art.
Nous nous intéressons, dans cet article, à la détection d’opinions dans la langue arabe. Ces dernières années, l’utilisation de l’apprentissage profond a amélioré des performances de nombreux systèmes automatiques dans une grande variété de domaines (analyse d’images, reconnaissance de la parole, traduction automatique, . . .) et également celui de l’analyse d’opinions en anglais. Ainsi, nous avons étudié l’apport de deux architectures (CNN et LSTM) dans notre cadre spécifique. Nous avons également testé et comparé plusieurs types de représentations continues de mots (embeddings) disponibles en langue arabe, qui ont permis d’obtenir de bons résultats. Nous avons analysé les erreurs de notre système et la pertinence de ces embeddings. Cette analyse mène à plusieurs perspectives intéressantes de travail, au sujet notamment de la constitution automatique de ressources expert et d’une construction pertinente des embeddings spécifiques à la tâche d’analyse d’opinions.
Le projet PASTEL étudie l’acceptabilité et l’utilisabilité des transcriptions automatiques dans le cadre d’enseignements magistraux. Il s’agit d’outiller les apprenants pour enrichir de manière synchrone et automatique les informations auxquelles ils peuvent avoir accès durant la séance. Cet enrichissement s’appuie sur des traitements automatiques du langage naturel effectués sur les transcriptions automatiques. Nous présentons dans cet article un travail portant sur l’annotation d’enregistrements de cours magistraux enregistrés dans le cadre du projet CominOpenCourseware. Ces annotations visent à effectuer des expériences de transcription automatique, segmentation thématique, appariement automatique en temps réel avec des ressources externes... Ce corpus comprend plus de neuf heures de parole annotées. Nous présentons également des expériences préliminaires réalisées pour évaluer l’adaptation automatique de notre système de reconnaissance de la parole.
Dialectal Arabic (DA) is significantly different from the Arabic language taught in schools and used in written communication and formal speech (broadcast news, religion, politics, etc.). There are many existing researches in the field of Arabic language Sentiment Analysis (SA); however, they are generally restricted to Modern Standard Arabic (MSA) or some dialects of economic or political interest. In this paper we are interested in the SA of the Tunisian Dialect. We utilize Machine Learning techniques to determine the polarity of comments written in Tunisian Dialect. First, we evaluate the SA systems performances with models trained using freely available MSA and Multi-dialectal data sets. We then collect and annotate a Tunisian Dialect corpus of 17.000 comments from Facebook. This corpus allows us a significant accuracy improvement compared to the best model trained on other Arabic dialects or MSA data. We believe that this first freely available corpus will be valuable to researchers working in the field of Tunisian Sentiment Analysis and similar areas.
Word embeddings have been successfully used in several natural language processing tasks (NLP) and speech processing. Different approaches have been introduced to calculate word embeddings through neural networks. In the literature, many studies focused on word embedding evaluation, but for our knowledge, there are still some gaps. This paper presents a study focusing on a rigorous comparison of the performances of different kinds of word embeddings. These performances are evaluated on different NLP and linguistic tasks, while all the word embeddings are estimated on the same training data using the same vocabulary, the same number of dimensions, and other similar characteristics. The evaluation results reported in this paper match those in the literature, since they point out that the improvements achieved by a word embedding in one task are not consistently observed across all tasks. For that reason, this paper investigates and evaluates approaches to combine word embeddings in order to take advantage of their complementarity, and to look for the effective word embeddings that can achieve good performances on all tasks. As a conclusion, this paper provides new perceptions of intrinsic qualities of the famous word embedding families, which can be different from the ones provided by works previously published in the scientific literature.
In this article, we present the RATP-DECODA Corpus which is composed by a set of 67 hours of speech from telephone conversations of a Customer Care Service (CCS). This corpus is already available on line at http://sldr.org/sldr000847/fr in its first version. However, many enhancements have been made in order to allow the development of automatic techniques to transcript conversations and to capture their meaning. These enhancements fall into two categories: firstly, we have increased the size of the corpus with manual transcriptions from a new operational day; secondly we have added new linguistic annotations to the whole corpus (either manually or through an automatic processing) in order to perform various linguistic tasks from syntactic and semantic parsing to dialog act tagging and dialog summarization.
L’étude présentée dans cet article améliore une méthode récemment proposée pour l’adaptation de modèles acoustiques markoviens couplés à un réseau de neurones profond (DNN-HMM). Cette méthode d’adaptation utilise des paramètres acoustiques dérivés de mixtures de modèles Gaussiens (GMM-derived features, GMMD ). L’amélioration provient de l’emploi de scores et de mesures de confiance calculés à partir de graphes construits dans le cadre d’un algorithme d’adaptation conventionnel dit de maximum a posteriori (MAP). Une version modifiée de l’adaptation MAP est appliquée sur le modèle GMM auxiliaire utilisé dans une procédure d’apprentissage adaptatif au locuteur (speaker adaptative training, SAT) lors de l’apprentissage du DNN. Des expériences menées sur le corpus Wall Street Journal (WSJ0) montrent que la technique d’adaptation non supervisée proposée dans cet article permet une réduction relative de 8, 4% du taux d’erreurs sur les mots (WER), par rapport aux résultats obtenus avec des modèles DNN-HMM indépendants du locuteur utilisant des paramètres acoustiques plus conventionnels.
L’étude porte sur l’apport d’un réseau de neurones récurrent (Recurrent Neural Network RNN) bidirectionnel encodeur/décodeur avec mécanisme d’attention pour une tâche de compréhension de la parole. Les premières expériences faites sur le corpus ATIS confirment la qualité du système RNN état de l’art utilisé pour cet article, en comparant les résultats obtenus à ceux récemment publiés dans la littérature. Des expériences supplémentaires montrent que les RNNs avec mécanisme d’attention obtiennent de meilleures performances que les RNNs récemment proposés pour la tâche d’étiquetage en concepts sémantiques. Sur le corpus MEDIA, un corpus français état de l’art pour la compréhension dédié à la réservation d’hôtel et aux informations touristiques, les expériences montrent qu’un RNN bidirectionnel atteint une f-mesure de 79,51 tandis que le même système intégrant le mécanisme d’attention permet d’atteindre une f-mesure de 80,27.
Récemment, l’utilisation des représentations continues de mots a connu beaucoup de succès dans plusieurs tâches de traitement du langage naturel. Dans cet article, nous proposons d’étudier leur utilisation dans une architecture neuronale pour la tâche de détection des erreurs au sein de transcriptions automatiques de la parole. Nous avons également expérimenté et évalué l’utilisation de paramètres prosodiques en suppléments des paramètres classiques (lexicaux, syntaxiques, . . .). La principale contribution de cet article porte sur la combinaison de différentes représentations continues de mots : plusieurs approches de combinaison sont proposées et évaluées afin de tirer profit de leurs complémentarités. Les expériences sont effectuées sur des transcriptions automatiques du corpus ETAPE générées par le système de reconnaissance automatique du LIUM. Les résultats obtenus sont meilleurs que ceux d’un système état de l’art basé sur les champs aléatoires conditionnels. Pour terminer, nous montrons que la mesure de confiance produite est particulièrement bien calibrée selon une évaluation en terme d’Entropie Croisée Normalisée (NCE).
Les travaux présentés portent sur l’extraction automatique d’unités sémantiques et l’évaluation de leur pertinence pour des conversations téléphoniques. Le corpus utilisé est le corpus français DECODA. L’objectif de la tâche est de permettre l’étiquetage automatique en thème de chaque conversation. Compte tenu du caractère spontané de ce type de conversations et de la taille du corpus, nous proposons de recourir à une stratégie semi-supervisée fondée sur la construction d’une ontologie et d’un apprentissage actif simple : un annotateur humain analyse non seulement les listes d’unités sémantiques candidates menant au thème mais étudie également une petite quantité de conversations. La pertinence de la relation unissant les unités sémantiques conservées, le sous-thème issu de l’ontologie et le thème annoté est évaluée par un DNN, prenant en compte une représentation vectorielle du document. L’intégration des unités sémantiques retenues dans le processus de classification en thème améliore les performances.
Dans cet article, nous nous intéressons au titrage automatique des segments issus de la segmentation thématique de journaux télévisés. Nous proposons d’associer un segment à un article de presse écrite collecté le jour même de la diffusion du journal. La tâche consiste à apparier un segment à un article de presse à l’aide d’une mesure de similarité. Cette approche soulève plusieurs problèmes, comme la sélection des articles candidats, une bonne représentation du segment et des articles, le choix d’une mesure de similarité robuste aux imprécisions de la segmentation. Des expériences sont menées sur un corpus varié de journaux télévisés français collectés pendant une semaine, conjointement avec des articles aspirés à partir de la page d’accueil de Google Actualités. Nous introduisons une métrique d’évaluation reflétant la qualité de la segmentation, du titrage ainsi que la qualité conjointe de la segmentation et du titrage. L’approche donne de bonnes performances et se révèle robuste à la segmentation thématique.
In this paper, we present improvements made to the TED-LIUM corpus we released in 2012. These enhancements fall into two categories. First, we describe how we filtered publicly available monolingual data and used it to estimate well-suited language models (LMs), using open-source tools. Then, we describe the process of selection we applied to new acoustic data from TED talks, providing additions to our previously released corpus. Finally, we report some experiments we made around these improvements.
In this paper we describe an effort to create a corpus and phonetic dictionary for Tunisian Arabic Automatic Speech Recognition (ASR). The corpus, named TARIC (Tunisian Arabic Railway Interaction Corpus) has a collection of audio recordings and transcriptions from dialogues in the Tunisian Railway Transport Network. The phonetic (or pronunciation) dictionary is an important ASR component that serves as an intermediary between acoustic models and language models in ASR systems. The method proposed in this paper, to automatically generate a phonetic dictionary, is rule based. For that reason, we define a set of pronunciation rules and a lexicon of exceptions. To determine the performance of our phonetic rules, we chose to evaluate our pronunciation dictionary on two types of corpora. The word error rate of word grapheme-to-phoneme mapping is around 9%.
This paper describes the Spoken Language Translation system developed by the LIUM for the IWSLT 2014 evaluation campaign. We participated in two of the proposed tasks: (i) the Automatic Speech Recognition task (ASR) in two languages, Italian with the Vecsys company, and English alone, (ii) the English to French Spoken Language Translation task (SLT). We present the approaches and specificities found in our systems, as well as the results from the evaluation campaign.
This paper presents the corpus developed by the LIUM for Automatic Speech Recognition (ASR), based on the TED Talks. This corpus was built during the IWSLT 2011 Evaluation Campaign, and is composed of 118 hours of speech with its accompanying automatically aligned transcripts. We describe the content of the corpus, how the data was collected and processed, how it will be publicly available and how we built an ASR system using this data leading to a WER score of 17.4 %. The official results we obtained at the IWSLT 2011 evaluation campaign are also discussed.
The PORTMEDIA project is intended to develop new corpora for the evaluation of spoken language understanding systems. The newly collected data are in the field of human-machine dialogue systems for tourist information in French in line with the MEDIA corpus. Transcriptions and semantic annotations, obtained by low-cost procedures, are provided to allow a thorough evaluation of the systems' capabilities in terms of robustness and portability across languages and domains. A new test set with some adaptation data is prepared for each case: in Italian as an example of a new language, for ticket reservation as an example of a new domain. Finally the work is complemented by the proposition of a new high level semantic annotation scheme well-suited to dialogue data.
This paper describes the three systems developed by the LIUM for the IWSLT 2011 evaluation campaign. We participated in three of the proposed tasks, namely the Automatic Speech Recognition task (ASR), the ASR system combination task (ASR_SC) and the Spoken Language Translation task (SLT), since these tasks are all related to speech translation. We present the approaches and specificities we developed on each task.
This paper presents the EPAC corpus which is composed by a set of 100 hours of conversational speech manually transcribed and by the outputs of automatic tools (automatic segmentation, transcription, POS tagging, etc.) applied on the entire French ESTER 1 audio corpus: this concerns about 1700 hours of audio recordings from radiophonic shows. This corpus was built during the EPAC project funded by the French Research Agency (ANR) from 2007 to 2010. This corpus increases significantly the amount of French manually transcribed audio recordings easily available and it is now included as a part of the ESTER 1 corpus in the ELRA catalog without additional cost. By providing a large set of automatic outputs of speech processing tools, the EPAC corpus should be useful to researchers who want to work on such data without having to develop and deal with such tools. These automatic annotations are various: segmentation and speaker diarization, one-best hypotheses from the LIUM automatic speech recognition system with confidence measures, but also word-lattices and confusion networks, named entities, part-of-speech tags, chunks, etc. The 100 hours of speech manually transcribed were split into three data sets in order to get an official training corpus, an official development corpus and an official test corpus. These data sets were used to develop and to evaluate some automatic tools which have been used to process the 1700 hours of audio recording. For example, on the EPAC test data set our ASR system yields a word error rate equals to 17.25%.
This paper describes the two systems developed by the LIUM laboratory for the 2010 IWSLT evaluation campaign. We participated to the new English to French TALK task. We developed two systems, one for each evaluation condition, both being statistical phrase-based systems using the the Moses toolkit. Several approaches were investigated.
This paper describes the systems developed by the LIUM laboratory for the 2009 IWSLT evaluation. We participated in the Arabic and Chinese to English BTEC tasks. We developed three different systems: a statistical phrase-based system using the Moses toolkit, an Statistical Post-Editing system and a hierarchical phrase-based system based on Joshua. A continuous space language model was deployed to improve the modeling of the target language. These systems are combined by a confusion network based approach.
Large vocabulary automatic speech recognition (ASR) technologies perform well in known, controlled contexts. However recognition of proper nouns is commonly considered as a difficult task. Accurate phonetic transcription of a proper noun is difficult to obtain, although it can be one of the most important resources for a recognition system. In this article, we propose methods of automatic phonetic transcription applied to proper nouns. The methods are based on combinations of the rule-based phonetic transcription generator LIA_PHON and an acoustic-phonetic decoding system. On the ESTER corpus, we observed that the combined systems obtain better results than our reference system (LIA_PHON). The WER (Word Error Rate) decreased on segments of speech containing proper nouns, without affecting negatively the results on the rest of the corpus. On the same corpus, the Proper Noun Error Rate (PNER, which is a WER computed on proper nouns only), decreased with our new system.
Our paper focuses on the gain which can be achieved on human transcription of spontaneous and prepared speech, by using the assistance of an ASR system. This experiment has shown interesting results, first about the duration of the transcription task itself: even with the combination of prepared speech + ASR, an experimented annotator needs approximately 4 hours to transcribe 1 hours of audio data. Then, using an ASR system is mostly time-saving, although this gain is much more significant on prepared speech: assisted transcriptions are up to 4 times faster than manual ones. This ratio falls to 2 with spontaneous speech, because of ASR limits for these data. Detailed results reveal interesting correlations between the transcription task and phenomena such as Word Error Rate, telephonic or non-native speech turns, the number of fillers or propers nouns. The latter make spelling correction very time-consuming with prepared speech because of their frequency. As a consequence, watching for low averages of proper nouns may be a way to detect spontaneous speech.
This paper describes the system developed by the LIUM laboratory for the 2008 IWSLT evaluation. We only participated in the Arabic/English BTEC task. We developed a statistical phrase-based system using the Moses toolkit and SYSTRAN’s rule-based translation system to perform a morphological decomposition of the Arabic words. A continuous space language model was deployed to improve the modeling of the target language. Both approaches achieved significant improvements in the BLEU score. The system achieves a score of 49.4 on the test set of the 2008 IWSLT evaluation.
This work adresses the use of confidence measures for extracting well recognized words with very low error rate from automatically transcribed segments in a unsupervised way. We present and compare several confidence measures and propose a method to merge them into a new one. We study its capabilities on extracting correct recognized word-segments compared to the amount of rejected words. We apply this fusion measure to select audio segments composed of words with a high confidence score. These segments come from an automatic transcription of french broadcast news given by our speech recognition system based on the CMU Sphinx3.3 decoder. Injecting new data resulting from unsupervised treatments of raw audio recordings in the training corpus of acoustic models gives statistically significant improvement (95% confident interval) in terms of word error rate. Experiments have been carried out on the corpus used during ESTER, the french evaluation campaign.
Le cadre de cette étude concerne les systèmes de dialogue via le téléphone entre un serveur de données et un utilisateur. Nous nous intéresserons au cas de dialogues non contraints où l’utilisateur à toute liberté pour formuler ses requêtes. Généralement, le module de Reconnaissance Automatique de la Parole (RAP) de tels serveurs utilise un seul Modèle de Langage (ML) de type bigramme ou trigramme pour modéliser l’ensemble des interventions possibles de l’utilisateur. Ces ML sont appris sur des corpus de phrases retranscrites à partir de sessions entre le serveur et plusieurs utilisateurs. Nous proposons dans cette étude une méthode de segmentation de corpus d’apprentissage de dialogue utilisant une stratégie mixte basée à la fois sur des connaissances explicites mais aussi sur l’optimisation d’un critère statistique. Nous montrons qu’un gain en terme de perplexité et de taux d’erreurs/mot peut être constaté en utilisant un ensemble de sous modèles de langage issus de la segmentation plutôt qu’un modèle unique appris sur l’ensemble du corpus.