L’exploration automatisée de grands corpus permet d’analyser plus finement la relation entre motifs de variation phonétique synchronique et changements diachroniques : les erreurs dans les transcriptions automatiques sont riches d’enseignements sur la variation contextuelle en parole continue et sur les possibles mutations systémiques sur le point d’apparaître. Dès lors, il est intéressant de se pencher sur des phénomènes phonologiques largement attestés dans les langues en diachronie comme en synchronie pour établir leur émergence ou non dans des langues qui n’y sont pas encore sujettes. La présente étude propose donc d’utiliser l’alignement forcé avec variantes de prononciation pour observer les alternances de voisement en coda finale de mot dans deux langues romanes : le français et le roumain. Il sera mis en évidence, notamment, que voisement et dévoisement non-canoniques des codas françaises comme roumaines ne sont pas le fruit du hasard mais bien des instances de dévoisement final et d’assimilation régressive de trait laryngal, qu’il s’agisse de voisement ou de non-voisement.

Cette étude vise à proposer une méthode adaptée à l’étude de divers phénomènes de variation dans les grands corpus utilisant l’alignement automatique de la parole. Cette méthode est appliquée pour étudier la réduction temporelle en français spontané. Nous proposons de qualifier la réduction temporelle comme la réalisation de suites de segments courts consécutifs. Environ 14% du corpus est considéré comme réduit. Les résultats de l’alignement montrent que ces zones impliquent le plus souvent plus d’un mot (81%), et que sinon, la position interne du mot est la plus concernée. Parmi les exemples de suites de mots les plus réduits, on trouve des locutions utilisées comme des marqueurs discursifs.

The present paper aims at providing a first study of lenition- and fortition-type phenomena in coda position in Romanian, a language that can be considered as less-resourced. Our data show that there are two contexts for devoicing in Romanian: before a voiceless obstruent, which means that there is regressive voicelessness assimilation in the language, and before pause, which means that there is a tendency towards final devoicing proper. The data also show that non-canonical voicing is an instance of voicing assimilation, as it is observed mainly before voiced consonants (voiced obstruents and sonorants alike). Two conclusions can be drawn from our analyses. First, from a phonetic point of view, the two devoicing phenomena exhibit the same behavior regarding place of articulation of the coda, while voicing assimilation displays the reverse tendency. In particular, alveolars, which tend to devoice the most, also voice the least. Second, the two assimilation processes have similarities that could distinguish them from final devoicing as such. Final devoicing seems to be sensitive to speech style and gender of the speaker, while assimilation processes do not. This may indicate that the two kinds of processes are phonologized at two different degrees in the language, assimilation being more accepted and generalized than final devoicing.

Computational Language Documentation attempts to make the most recent research in speech and language technologies available to linguists working on language preservation and documentation. In this paper, we pursue two main goals along these lines. The first is to improve upon a strong baseline for the unsupervised word discovery task on two very low-resource Bantu languages, taking advantage of the expertise of linguists on these particular languages. The second consists in exploring the Adaptor Grammar framework as a decision and prediction tool for linguists studying a new language. We experiment 162 grammar configurations for each language and show that using Adaptor Grammars for word segmentation enables us to test hypotheses about a language. Specializing a generic grammar with language specific knowledge leads to great improvements for the word discovery task, ultimately achieving a leap of about 30% token F-score from the results of a strong baseline.

La transcription automatique de la parole obtient aujourd’hui des performances élevées avec des taux d’erreur qui tombent facilement en dessous de 10% pour une parole journalistique. Cependant, pour des conversations plus libres, ils stagnent souvent autour de 20–30%. En français, une grande partie des erreurs sont dues à des confusions entre homophones n’impliquant pas les niveaux acousticophonétique et phonologique. Cependant, de nombreuses erreurs peuvent s’expliquer par des variantes de productions non prévues par le système. Afin de mieux comprendre quels processus phonologiques pourraient expliquer ces variantes spécifiques de la parole spontanée, nous proposons une analyse des erreurs en comparant prononciations attendue (référence) et reconnue (hypothèse) via un alignement phonétique par programmation dynamique. Les distances locales entre paires de phonèmes appariés correspondent au nombre de traits phonétiques disjoints. Nos analyses permettent d’identifier les traits phonétiques les plus fréquemment impliqués dans les erreurs et donnent des pistes pour des interprétations phonologiques.

Quelles sont les caractéristiques acoustiques et articulatoires des voyelles parlées et chantées du Cantu in Paghjella (polyphonie corse à trois voix), en fonction du chanteur, de la voyelle et de la fréquence fondamentale ? L’analyse acoustique des quatre premiers formants de la parole au chant et celle des mouvements articulatoires lingual et labial, montrent généralement (i) une significative augmentation de F1 avec abaissement lingual mais fermeture labiale, en lien avec une corrélation entre F0 et F1 ; (ii) une baisse de F2 pour les voyelles antérieures, une postériorisation linguale et un recul de l’ombre hyoïdienne uniquement pour le bassu ; (iii) une nette augmentation de F3 et F4 surtout chez le bassu ; (iv) une augmentation du Singing Power Ratio surtout chez les bassu et secunda. Ses valeurs sont toutefois inférieures à celles de chanteurs lyriques, et ne correspondant pas comme ces derniers à un rapprochement de F3 et F4.

Le rôle du contexte est connu dans la réalisation ou non du schwa en français. Deux grands corpus oraux de parole journalistique (ETAPE) et de parole familière (NCCFr), dans lesquels la realisation de schwa est déterminée à partir d’un alignement automatique, ont été utilisés pour examiner la contribution du contexte au sein du mot contenant schwa (lexical) vs. au travers de la frontière avec le mot précédent (post-lexical). Nos résultats montrent l’importance du contexte pré-frontière dans l’explication de la chute du schwa dans la première syllabe d’un mot polysyllabique en parole spontanée. Si le mot précédant se termine par une consonne, nous pouvons faire appel à la loi des trois consonnes et au principe de sonorité pour expliquer des différences de comportement en fonction de la nature des consonnes en contact.

Les apprenants français de l’allemand ont des difficultés à produire la fricative palatale sourde allemande /ç/ (Ich-Laut) et ont tendance à la remplacer par la fricative post-alvéolaire /S/. Nous nous demandons si avec des mesures acoustiques ces imprécisions de production peuvent être quantifiées d’une manière plus objective. Deux mesures acoustiques ont été examinées afin de distinguer au mieux /S/ et /ç/ dans un contexte VC en position finale de mot dans des productions de locuteurs germanophones natifs. Elles servent ensuite à quantifier les difficultés de production des apprenants français. 285 tokens de 20 locuteurs natifs et 20 locuteurs L2 ont été analysés. Les mesures appliquées sont le centre de gravité spectral et des rapports d’intensité par bande de fréquence. Sur les productions de locuteurs natifs, les résultats montrent que la mesure la plus fiable pour distinguer acoustiquement /S/ et /ç/ est le ratio d’intensité entre fréquences hautes (4-7 kHz) et basses (1-4 kHz). Les mesures confirment également les difficultés de production des locuteurs natifs français.

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).

The French Learners Audio Corpus of German Speech (FLACGS) was created to compare German speech production of German native speakers (GG) and French learners of German (FG) across three speech production tasks of increasing production complexity: repetition, reading and picture description. 40 speakers, 20 GG and 20 FG performed each of the three tasks, which in total leads to approximately 7h of speech. The corpus was manually transcribed and automatically aligned. Analysis that can be performed on this type of corpus are for instance segmental differences in the speech production of L2 learners compared to native speakers. We chose the realization of the velar nasal consonant engma. In spoken French, engma does not appear in a VCV context which leads to production difficulties in FG. With increasing speech production complexity (reading and picture description), engma is realized as engma + plosive by FG in over 50% of the cases. The results of a two way ANOVA with unequal sample sizes on the durations of the different realizations of engma indicate that duration is a reliable factor to distinguish between engma and engma + plosive in FG productions compared to the engma productions in GG in a VCV context. The FLACGS corpus allows to study L2 production and perception.

Luxembourgish, embedded in a multilingual context on the divide between Romance and Germanic cultures, remains one of Europe’s under-described languages. This is due to the fact that the written production remains relatively low, and linguistic knowledge and resources, such as lexica and pronunciation dictionaries, are sparse. The speakers or writers will frequently switch between Luxembourgish, German, and French, on a per-sentence basis, as well as on a sub-sentence level. In order to build resources like lexicons, and especially pronunciation lexicons, or language models needed for natural language processing tasks such as automatic speech recognition, language used in text corpora should be identified. In this paper, we present the design of a manually annotated corpus of mixed language sentences as well as the tools used to select these sentences. This corpus of difficult sentences was used to test a word-based language identification system. This language identification system was used to select textual data extracted from the web, in order to build a lexicon and language models. This lexicon and language model were used in an Automatic Speech Recognition system for the Luxembourgish language which obtain a 25\% WER on the Quaero development data.

This paper is concerned with human assessments of the severity of errors in ASR outputs. We did not design any guidelines so that each annotator involved in the study could consider the “seriousness” of an ASR error using their own scientific background. Eight human annotators were involved in an annotation task on three distinct corpora, one of the corpora being annotated twice, hiding this annotation in duplicate to the annotators. None of the computed results (inter-annotator agreement, edit distance, majority annotation) allow any strong correlation between the considered criteria and the level of seriousness to be shown, which underlines the difficulty for a human to determine whether a ASR error is serious or not.

This paper addresses the question of hierarchical named entity evaluation. In particular, we focus on metrics to deal with complex named entity structures as those introduced within the QUAERO project. The intended goal is to propose a smart way of evaluating partially correctly detected complex entities, beyond the scope of traditional metrics. None of the existing metrics are fully adequate to evaluate the proposed QUAERO task involving entity detection, classification and decomposition. We are discussing the strong and weak points of the existing metrics. We then introduce a new metric, the Entity Tree Error Rate (ETER), to evaluate hierarchical and structured named entity detection, classification and decomposition. The ETER metric builds upon the commonly accepted SER metric, but it takes the complex entity structure into account by measuring errors not only at the slot (or complex entity) level but also at a basic (atomic) entity level. We are comparing our new metric to the standard one using first some examples and then a set of real data selected from the ETAPE evaluation results.

It is well-known that human listeners significantly outperform machines when it comes to transcribing speech. This paper presents a progress report of the joint research in the automatic vs human speech transcription and of the perceptual experiments developed at LIMSI that aims to increase our understanding of automatic speech recognition errors. Two paradigms are described here in which human listeners are asked to transcribe speech segments containing words that are frequently misrecognized by the system. In particular, we sought to gain information about the impact of increased context to help humans disambiguate problematic lexical items, typically homophone or near-homophone words. The long-term aim of this research is to improve the modeling of ambiguous contexts so as to reduce automatic transcription errors.

Text and speech corpora for training a tale telling robot have been designed, recorded and annotated. The aim of these corpora is to study expressive storytelling behaviour, and to help in designing expressive prosodic and co-verbal variations for the artificial storyteller). A set of 89 children tales in French serves as a basis for this work. The tales annotation principles and scheme are described, together with the corpus description in terms of coverage and inter-annotator agreement. Automatic analysis of a new tale with the help of this corpus and machine learning is discussed. Metrics for evaluation of automatic annotation methods are discussed. A speech corpus of about 1 hour, with 12 tales has been recorded and aligned and annotated. This corpus is used for predicting expressive prosody in children tales, above the level of the sentence.

The national language of the Grand-Duchy of Luxembourg, Luxembourgish, has often been characterized as one of Europe's under-described and under-resourced languages. Because of a limited written production of Luxembourgish, poorly observed writing standardization (as compared to other languages such as English and French) and a large diversity of spoken varieties, the study of Luxembourgish poses many interesting challenges to automatic speech processing studies as well as to linguistic enquiries. In the present paper, we make use of large corpora to focus on typical writing and derived pronunciation variants in Luxembourgish, elicited by mobile -n deletion (hereafter shortened to MND). Using transcriptions from the House of Parliament debates and 10k words from news reports, we examine the reality of MND variants in written transcripts of speech. The goal of this study is manyfold: quantify the potential of variation due to MND in written Luxembourgish, check the mandatory status of the MND rule and discuss the arising problems for automatic spoken Luxembourgish processing.

The goal of this paper is to investigate French word segmentation strategies using phonemic and lexical transcriptions as well as prosodic and part-of-speech annotations. Average fundamental frequency (f0) profiles and phoneme duration profiles are measured using 13 hours of broadcast news speech to study prosodic regularities of French words. Some influential factors are taken into consideration for f0 and duration measurements: word syllable length, word-final schwa, part-of-speech. Results from average f0 profiles confirm word final syllable accentuation and from average duration profiles, we can observe long word final syllable length. Both are common tendencies in French. From noun phrase studies, results of average f0 profiles illustrate higher noun first syllable after determiner. Inter-vocalic duration profile results show long inter-vocalic duration between determiner vowel and preceding word vowel. These results reveal measurable cues contributing to word boundary location. Further studies will include more detailed within syllable f0 patterns, other speaking styles and languages.

This paper presents a preliminary analysis of the role of some discourse markers and the vocalic hesitation ""euh"" in a corpus of spoken human utterances collected with the Ritel system, an open domain and spoken dialog system. The frequency and contextual combinatory of classical discourse markers and of the vocalic hesitation have been studied. This analysis pointed out some specificity in terms of combinatory of the analyzed items. The classical discourse markers seem to help initiating larger discursive blocks both at initial and medial positions of the on-going turns. The vocalic hesitation stand also for marking the user's embarrassments and wish to close the dialog.

The performance of question answering system is evaluated through successive evaluations campaigns. A set of questions are given to the participating systems which are to find the correct answer in a collection of documents. The creation process of the questions may change from one evaluation to the next. This may entail an uncontroled question difficulty shift. For the QAst 2009 evaluation campaign, a new procedure was adopted to build the questions. Comparing results of QAst 2008 and QAst 2009 evaluations, a strong performance loss could be measured in 2009 for French and English, while the Spanish systems globally made progress. The measured loss might be related to this new way of elaborating questions. The general purpose of this paper is to propose a measure to calibrate the difficulty of a question set. In particular, a reasonable measure should output higher values for 2009 than for 2008. The proposed measure relies on a distance measure between the critical elements of a question and those of the associated correct answer. An increase of the proposed distance measure for French and English 2009 evaluations as compared to 2008 could be established. This increase correlates with the previously observed degraded performances. We conclude on the potential of this evaluation criterion: the importance of such a measure for the elaboration of new question corpora for questions answering systems and a tool to control the level of difficulty for successive evaluation campaigns.

In this paper, we investigate the acoustic properties of phonemes in three speaking styles: read speech, prepared speech and spontaneous speech. Our aim is to better understand why speech recognition systems still fails to achieve good performances on spontaneous speech. This work follows the work of Nakamura et al. on Japanese speaking styles, with the difference that we here focus on French. Using Nakamura's method, we use classical speech recognition features, MFCC, and try to represent the effects of the speaking styles on the spectral space. Two measurements are defined in order to represent the spectral space reduction and the spectral variance extension. Experiments are then carried on to investigate if indeed we find some differences between the three speaking styles using these measurements. We finally compare our results to those obtained by Nakamura on Japanese to see if the same phenomenon appears. We happen to find some cues, and it also seems that phone duration also plays an important role regarding spectral reduction, especially for spontaneous speech.

Looking for a better understanding of spontaneous speech-related phenomena and to improve automatic speech recognition (ASR), we present here a study on the relationship between the occurrence of overlapping speech segments and disfluencies (filled pauses, repetitions, revisions) in political interviews. First we present our data, and our overlap annotation scheme. We detail our choice of overlapping tags and our definition of disfluencies; the observed ratios of the different overlapping tags are examined, as well as their correlation with of the speaker role and propose two measures to characterise speakers’ interacting attitude: the attack/resist ratio and the attack density. We then study the relationship between the overlapping speech segments and the disfluencies in our corpus, before concluding on the perspectives that our experiments offer.

In the present contribution we start with an overview of the linguistic situation of Luxembourg. We then describe specificities of spoken and written Lëtzebuergesch, with respect to automatic speech processing. Multilingual code-switching and code-mixing, poor writing standardization as compared to languages such as English or French, a large diversity of spoken varieties, together with a limited written production of Lëtzebuergesch language contribute to pose many interesting challenges to automatic speech processing both for speech technologies and linguistic studies. Multilingual filtering has been investigated to sort out Luxembourgish from German and French. Word list coverage and language model perplexity results, using sibling resources collected from the Web, are presented. A phonemic inventory has been adopted for pronunciation dictionary development, a grapheme-phoneme tool has been developed and pronunciation research issues related to the multilingual context are highlighted. Results achieved in resource development allow to envision the realisation of an ASR system.

The present contribution aims at increasing our understanding of automatic speech recognition (ASR) errors involving frequent homophone or almost homophone words by confronting them to perceptual results. The long-term aim is to improve acoustic modelling of these items to reduce automatic transcription errors. A first question of interest addressed in this paper is whether homophone words such as “et” (and); and “est” (to be), for which ASR systems rely on language model weights, can be discriminated in a perceptual transcription test with similar n-gram constraints. A second question concerns the acoustic separability of the two homophone words using appropriate acoustic and prosodic attributes. The perceptual test reveals that even though automatic and perceptual errors correlate positively, human listeners deal with local ambiguity more efficiently than the ASR system in conditions which attempt to approximate the information available for decision for a 4-gram language model. The corresponding acoustic analysis shows that the two homophone words may be distinguished thanks to some relevant acoustic and prosodic attributes. A first experiment in automatic classification of the two words using data mining techniques highlights the role of the prosodic (duration and voicing) and contextual information (pauses co-occurrence) in distinguishing the two words. Current results, even though preliminary, suggests that new levels of information, so far unexplored in pronunciations’ modelling for ASR, may be considered in order to efficiently factorize the word variants observed in speech and to improve the automatic speech transcription.

The present study focuses on automatic processing of sibling resources of audio and written documents, such as available in audio archives or for parliament debates: written texts are close but not exact audio transcripts. Such resources deserve attention for several reasons: they represent an interesting testbed for studying differences between written and spoken material and they yield low cost resources for acoustic model training. When automatically transcribing the audio data, regions of agreement between automatic transcripts and written sources allow to transfer time-codes to the written documents: this may be helpful in an audio archive or audio information retrieval environment. Regions of disagreement can be automatically selected for further correction by human transcribers. This study makes use of 10 hours of French radio interview archives with corresponding press-oriented transcripts. The audio corpus has then been transcribed using the LIMSI speech recognizer resulting in automatic transcripts, exhibiting an average word error rate of 12%. 80% of the text corpus (with word chunks of at least five words) can be exactly aligned with the automatic transcripts of the audio data. The residual word error rate on these 80% is less than 1%.