SPONTANEOUS SPEECH RECOGNITION FOR ROMANIAN IN SPOKEN DIALOGUE SYSTEMS

Abstract

In this paper we present an attempt to develop a speech recognition module for the Romanian language in order to be used in a dialogue system. The main characteristics of such a dialogue system are first discussed. Further, we explain the design and acquisition of a spontaneous speech database for training the decoder: the design guidelines in developing the database, as well as several practical issues encountered, along with some triphones balancing statistics are pointed out. Then, the speech recognition architecture (based on components in the “Hidden Markov Modeling Toolkit ” – HTK) is described in detail, emphasizing the two aspects, training and decoding. In the next section, a discussion of several preliminary recognition results is provided, emphasizing current limitations and the need to significantly increase the size of the database. A set of conclusions and perspectives are offered at the end of the paper.