Section: Overall Objectives

Introduction

PAROLE is a joint project to Inria, CNRS, Lorraine University through the LORIA laboratory (UMR 7503). The purpose of our project is to automatically process speech signals to understand their meaning, and to analyze and enhance their acoustic structure. It inscribes within the view of offering efficient vocal technologies and necessitates works in analysis, perception and automatic recognition (ASR) of speech.

Our activities are structured in three topics:

Speech analysis and synthesis. Our works are concerned with automatic extraction and perception of acoustic and visual cues, acoustic-to-articulatory inversion and speech synthesis. These themes give rise to a number of ongoing and future applications especially in the domain of foreign language learning.

Enriched automatic speech recognition. Our works are concerned with stochastic models (HMM (Hidden Markov Models) and Bayesian networks), semi-supervised and smoothed training of these stochastic models, adaptation of a recognition system to important variabilities, and with enriching the output of speech recognition with higher-level information such as syntactic structure and punctuation marks. These topics give also rise to a number of ongoing and future applications: automatic transcription, speech/text alignment, audio indexing, keyword spotting, foreign language learning, dialog systems, vocal services...

Speech to Speech Translation and Langage Modeling. This axis concerns statistical machine translation. The objective is to translate speech from a source language to any target language. The main activity of the group which is in charge of this axis is to propose an alternative method to the classical five IBM's models. This activity should conduct to several applications: e-mail speech to text, translation of movie subtitles.

Our pluridisciplinary scientific culture combines works in phonetics, pattern recognition and artificial intelligence. This pluridisciplinarity turns out to be a decisive asset to address new research topics, particularly language learning that simultaneously require competences in automatic speech recognition and phonetics.

Our policy in terms of industrial partnership consists in favoring contracts that quite precisely fit our scientific objectives. We are involved in an ANR project about audiovisual speech synthesis, another about acoustic-to-articulatory inversion of speech (ARTIS), another about the processing of articulatory data (DOCVACIM) and in a national evaluation campaign of automatic speech recognition systems (ESTER). We also coordinated until January 2009 the 6th PCRD project ASPI about acoustic-to-articulatory inversion of speech, and the Rapsodis ARC until october 2009. Additionally, we are also participating to a number of regional projects.