Description:
<jats:p>A quick assessment of the current state-of-the-art in Speech Recognition and Synthesis allows to note that in spite of positive advances in the last years </jats:p>
<jats:p>1) no current recognition system is capable of processing with reliability really spontaneous continuous speech, </jats:p>
<jats:p>2) no current speech synthesis system provides speech which could be confused with natural speech, </jats:p>
<jats:p>3) in these two domains, for the last fifteen years, "knowledge-based" approaches have been less fruitful than "statistical training" approaches, </jats:p>
<jats:p>4) the best current systems merely copy "surface" speech phenomena, </jats:p>
<jats:p>5) such "imitation" has been essentially limited to the language performances observed in two kinds of task (as to synthesis, loud- speaking reading of written texts ; as to recognition, written transcriptions of read texts), underscoring the teleological dimension of "speech" activity (case of dialogue). </jats:p>
<jats:p>If we want to be able, some day, to communicate with a machine in a natural way, it seems crucial to us to consider, from now on, speech recognition and synthesis no longer as elementary, task-independent processes but as components of complete systems of human-machine communication by speech. That will make necessary to integrate closely speech production and recognition functionalities to the controlling and monitoring organ which, in prospect of achieving a given task, will run a reasoning that will allow to understand or generate a message, using various knowledge sources, optimally structured and managed. The association of deeper knowledge and more efficient modelling of language performances and behaviours (included learning mechanisms...) in various contexts and for various tasks seems to us a key factor of future advances in that domain. </jats:p>