Guy Perennou

Structuration des informations acoustiques dans le projet A.R.I.A.L.

Abstract The A.R.I.A.L. (Analyse et Reconnaissance des Informations Acoustiques et Linguistiques = Analysis and recognition of acoustic and linguistic data.) project uses acoustic and phonetic data coming from an ear model and structured in an ascending manner. The data are sometimes organized as features and sometimes as patterns and take into account context effects coded in the same way as instantaneous parameters. Our discussion will concern neither the features nor the pattern concepts, nor indeed their efficiency for recognition. What we shall do is to describe the A.R.I.A.L. project data basis. 1. 1. The acoustic data are organized hierarchically: on level 1: spectral data, intensity and pitch, on level 2: acoustic cues, on level 3: discrete cues (pseudo-features). 2. 2. Segmentation produces short, homogeneous segments (infra-phonemes) in whose all the cues take part in the implementation. A phonetic unit (phoneme, syllable, etc.) is described as a series of segments labelled Si. 3. 3. The knowledge basis is built up automatically without the help of an expert except for the initial manual segmentation of the sentences analysed. All the segments encountered make up an alphabet (Si). 3.1. Reduction of the alphabet by means of a metric routine. 3.2. Setting up of the base: Uk/ClSaSb—, Unit Uk in context Cl is a series of (Si) segments. It is then possible to define the classes of equivalence between series of segments, and this makes it possible (a) to give a single rewrite rule for Uk/Cl, (b) to study the oppositions between units. 3.3. Reduction of the data basis using a metric routine and dynamic programming.

Besoins lexicaux a la lumiere de l'analyse statistique du corpus de textes du projet BREF: le lexique BDLEX du francais ecrit et oral

In this paper, we describe lexical needs for spoken and written French surface processing, like automatic text correction, speech recognition and synthesis.We present statistical observations made on a vocabulary compiled from real texts like articles. These texts have been used for building a recorded speech database called BREF. Developed by the Limsi, within the research group GDR-PRC CHM (Groupe De Recherche - Programme de Recherches Concertees, Communication Homme-Machine --- Research Group - Concerted Research Program, Man Machine Communication), this database is intended for dictation machine development and assessment.In this study, the informations available in our lexical database BDLEX (Base de Donnees LEXicales - Lexical Database) are used as reference materials. Belonging to the same research group than BREF, BDLEX has been developed for spoken and written French. Its purpose is to create, organize and provide lexical materials intended for automatic speech and text processing.Lexical covering takes an important part in such system assessment. Our first purpose is to value the rate of lexical covering that a 50, 000 word lexicon can reach.By comparison between the vocabulary provided (LexBref, composed of 84, 900 items, mainly distinct inflected forms) and the forms generated from BDLEX, we obtain about 62% of known forms, taking in account some acronyms and abbreviations.Then, we approach the unexpected word question looking into the 38% of left forms. Among them we can find numeration, neologisms, foreign words and proper names, as well as other acronyms and abbreviations. So, to obtain a large text covering, a lexical component must take in account all these kinds of words and must be fault tolerant, particularly with typographic faults.Last, we give a general description of the BDLEX project, specially of its lexical content. We describe some lexical data recently inserted in BDLEX according to the observations made on real texts. It concerns more particularly the lexical item representation using phonograms (i.e. letters/sounds associations), informations about acronyms and abbreviations as well as morphological knowledge about derivative words. We also present a set of linguistic tools connected to BDLEX and working on the phonological, orthographical and morphosyntactical levels.

Specifications pour un systeme generateur de modeles de decodage phonetique

Abstract The complexity of phonetic decoding presents automatic speech recognition (ASR) systems with a number of imperfectly solved problems which concern both phoneticians and psycho-linguists. An interesting contribution may be expected from the use of computers which make wide-ranging experiments and the easy introduction of phonetic data possible. But purely declarative data will be insufficient for the ‘competence’ they confer. Everything affecting performance i.e. here, decoding strategies, must also be added. An expert system largely fulfils the above requirements but presupposes that work is advanced enough for proven strategies and an efficient representation of the data basis to be set up. As in the case of EMYCIN and HEARSAY III, we prefer to set up a systemgenerating system which, among other things, allows the experimental model studied in ARIAL II to be finalized. This paper will give the general specifications of a generating system capable of: (a) creating generalized transduction networks integrating: simple or procedural declarative knowledge, and metaknowledge; (b) checking the decoding of acoustico-phonetic data guided by the above networks and instructions included in the questions asked, such as: verifying a syllable or family of syllables (or another type of entity), or unrestricted ascending search.